Research Michigan Center Retirement

advertisement
Michigan
Retirement
Research
University of
Working Paper
WP 2003-056
Center
Labor Supply Flexibility and Portfolio Choice:
An Empirical Analysis
Hugo Benítez-Silva
MR
RC
Project #: UM03-09
“Labor Supply Flexibility and Portfolio Choice:
An Empirical Analysis”
Hugo Benítez-Silva
SUNY at Stony Brook
June 2003
Michigan Retirement Research Center
University of Michigan
P.O. Box 1248
Ann Arbor, MI 48104
Acknowledgements
This work was supported by a grant from the Social Security Administration through the
Michigan Retirement Research Center (Grant # 10-P-98358-5). The opinions and
conclusions are solely those of the authors and should not be considered as representing
the opinions or policy of the Social Security Administration or any agency of the Federal
Government.
Regents of the University of Michigan
David A. Brandon, Ann Arbor; Laurence B. Deitch, Bingham Farms; Olivia P. Maynard, Goodrich;
Rebecca McGowan, Ann Arbor; Andrea Fischer Newman, Ann Arbor; Andrew C. Richner, Grosse Pointe
Park; S. Martin Taylor, Gross Pointe Farms; Katherine E. White, Ann Arbor; Mary Sue Coleman, ex
officio
Labor Supply Flexibility and Portfolio Choice:
An Empirical Analysis
Hugo Benítez-Silva
Abstract
This paper uses panel data from the Health and Retirement Study to estimate the
relationship between measures of labor supply flexibility and portfolio-choice decisions
by utility-maximizing individuals. Seminal research on portfolio decisions over the lifecycle, and recent research on stochastic dynamic programming models with endogenous
labor supply and savings decisions suggest that, other things equal, individuals with more
labor supply flexibility are likely to invest more in risky assets, regardless of their age,
because of the insurance component that flexible labor supply provides. After controlling
for panel sample selection and unobserved heterogeneity I find that labor supply
flexibility leads to holding between 12% and 14% more wealth in stocks.
Authors’ Acknowledgements
I am grateful to Sofia Cheidvasser, Thomas Knaus, Lorenzo Forni, Michael Hurd, Arie
Kapteyn, Deb Dwyer, Mark Montgomery, Kevin Austin, Joseph Tracy, Bart Hobijn,
Katja Seim, Caroline Austin, and Anna Vellv´e-Torras for their comments, and the
participants of the RAND Aging Workshop, the University of Colorado at Boulder
Seminar Series, the North American Meetings of the Econometric Society at UCLA, the
2002 Meetings of the Society for Economic Dynamics at NYU, the SOLE 2001 Meetings
in Austin, and the Yale Summer Microeconomics Seminars, for insightful discussions.
Frank Heiland, Wayne-Roy Gayle, and Huan Ni provided superb research assistance. I
am also grateful for the financial support of the TIAA-CREF Institute (2002-2003) and
the Michigan Retirement Research Center (UM03-09). The Cowles Foundation for
Research in Economics through a Carl Arvid Anderson Dissertation Fellowship, and the
John Perry Miller Fund supported me during the early stages of this project. I also want
to thank the Department of Economics at Universitat Aut`onoma de Barcelona, and the
Economics Department at the University of Maryland for their hospitality during the
Summer of 2002. I bear sole responsibility for any remaining errors.
1 Introduction
In this paper I empirically analyze the relationship between measures of labor supply flexibility and
portfolio-choice decisions by utility-maximizing individuals using panel data from the Health and Retirement Study (HRS). Once I control for unobserved heterogeneity and selection into the panel, I find that
individuals with labor supply flexibility hold on average between 12% and 14% more wealth in the form of
stocks in their portfolios. The selection into the market for stocks is of independent interest but also a key ingredient of the overall analysis. This is one of the first empirical studies that follows the work of Nijman and
Verbeek (1992), Verbeek and Nijman (1992), Vella (1998), Vella and Verbeek (1999), and the recommendations in Jensen, Rosholm and Verner (2001) to control for sample selection in panel data models. I find
that the additional computational burden of estimating these more sophisticated specifications is justified by
the results that show that selection into the market for stocks has an individual specific component (which
reflects stock ownership differences across individuals) which positively affects stock holdings, but also an
idiosyncratic component (which reflects the stock ownership differences over time for a given individual)
which negatively affects stock holdings. Traditional cross-sectional based selection models seem to only
capture the latter.
The connection between labor supply flexibility and portfolio choice has been largely overlooked by
labor and financial economists. The labor literature has focused its attention on participation decisions and
on consumption/saving behavior, and only seldom has modeled both together. On the finance side, labor
income has been acknowledged as an important source of balances, and sometimes even included in the
models, but labor supply has almost always been taken as exogenous. 1 One possible reason for the lack
of attention on this issue is the absence of models that show the importance of this connection. Without
theoretical models to guide researchers, empirical work on the issue has also been limited.
However, it is difficult to ignore the interactions between these two issues. Most individuals would
acknowledge a relationship between the amount of work they decide to do and their financial wealth and its
evolution. It is logical to think of labor income as a type of insurance against bad investment outcomes, and
it is also reasonable to believe that labor flexibility ex-post can have an effect in the investment decisions
ex-ante. At the end of the day whether there is a relationship between labor supply flexibility and portfolio
allocation becomes an empirical question, but before I test this relationship it makes sense to mention some
models that have emphasized this connection.
Bodie and Samuelson (1989) and Bodie, Merton, and Samuelson (1992) make the connection using
numerical examples using a two period model and also the continuous time portfolio theory popularized
1 See Killingsworth (1983) for a classical treatment of labor supply models, Heckman (1974) for a model of consumption/saving
with endogenous labor supply, and Merton (1990) for an overview of models of portfolio allocation.
1
by Merton (1969, 1990). They show that if labor supply flexibility is allowed in a continuous-time life
cycle model of consumption/saving and leisure choices, the holdings of risky assets would be positively
affected. In other words, individuals with more labor supply flexibility are predicted to hold more risky
assets. Their models emphasize the importance of considering human capital along with traditional financial
wealth as part of an individual’s portfolio. The authors, in an illuminating concluding section, conjecture
that the relationship between labor supply flexibility and holdings of risky assets is likely to be positive in
an empirical study that could adequately measure the relevant variables in the model.
More recently, a number of studies have approached the problem of solving a stochastic dynamic lifecycle model of consumption-saving allowing for endogenous choices of labor supply. 2 Among these studies,
Benı́tez-Silva (2002) also endogenizes the annuity decision over the life cycle. His results show that when
labor supply is endogenous individuals prefer to invest more in the risky asset than in the less risky annuity.
He relaxes several assumptions made in the continuous-time framework, allowing for lifetime and capital
uncertainty, bequests, non-separability between consumption and leisure in the utility function, endogenous
retirement, and the presence of Social Security. 3
My research complements the empirical efforts to disentangle what has been called the “equity premium
puzzle at the micro level,” the question as to why individuals’ portfolios are relatively low on risky assets. 4
The puzzle at the macro level has received an impressive amount of attention. The task of explaining the
large gap between the historical returns of risky and riskless assets has been a much more challenging enterprise than anyone could have expected, and as Kocherlakota (1996) explains, it shows the large gaps in
our understanding of the macroeconomy. 5 The microeconomic side of the puzzle also brings home our difficulties in explaining the behavior of economic agents without resorting to implausible assumptions or the
substantial modification of theories widely believed to be supported by microeconomic and macroeconomic
empirical evidence.6
Many explanations have been put forth to understand why individuals do not hold risky assets or they
hold small amounts of them on average. A relatively new literature has focused on the presence of uninsurable (labor) income risk and it presents a precautionary argument to justify the low holdings of risky
2
See for example Flodén (1998), Low (1998, and 1999), French (2000), Rust, Buchinsky, and Benı́tez-Silva (2002), and
Benı́tez-Silva (2002).
3 This model complements and extends the work of Yaari (1965), Friedman and Warshawsky (1990), and Mitchell et al. (1999).
4 The proportion of stocks on the average portfolio of American households has increased from around 5% to almost 12%
during the 1990s, but researchers have mostly blamed the impressive returns on stocks (more than twice the historical average)
that investors enjoyed during the last decade. See Tracy and Schneider (2001) for an overview. See Guiso, Haliassos, and Jappelli
(2002) for an up-to-date overview of research on household portfolios.
5 See also Cochrane (1997) for an interesting presentation of the facts and theories.
6 Campbell et al. (2001) also emphasize the importance of understanding the microeconomic problem. They solve an infinite
horizon portfolio choice problem with a time-varying equity premium. In their model labor supply does not play any role and there
is no risky labor income.
2
assets. Koo (1999), in a finite horizon discrete time setting, shows that liquidity constraints and uninsurable
income risk reduce consumption and investment in the risky asset. Heaton and Lucas (2000a) calibrate a
infinitely-lived agent dynamic portfolio choice model where background risk is shown to have an effect on
portfolio allocations, and Campbell and Cocco (2002) solve a finite horizon model that allows for risky labor income and borrowing constraints, in order to analyze the optimal mortgage choice. In all these models
labor supply does not play any role. 7 On the empirical side, most studies find some support for this argument, but the evidence seems to be weak. For example, Guiso et al. (1996), using Italian data, find some
evidence to support the conjecture that higher labor income risk is associated with lower holdings of risky
assets, Arrondel (2000) using French data finds weak support for the importance of labor income risk, and
Hochguertel (2003) and Alessie, Hochguertel, and van Soest (2002) using panel data from the Netherlands
find that after controlling for unobserved heterogeneity, the link between labor income risk and the holdings
of risky assets is very weak.
Other authors have emphasized other explanations: Vissing-Jorgenson (1999), and Bertaut and StarrMcCluer (2002) point out the importance of transaction and participation costs, and Lusardi (1999) focuses
on the importance of information costs. Yet others, like Fratantoni (1998), using the Survey of Consumer
Finances, and Robst et al. (1999) using the PSID, introduce the issue of housing, given that it can be
considered a risky investment, affected by labor income uncertainty, and that affects the holding of other
potentially risky assets. On the other hand, Heaton and Lucas (2000b) have presented evidence that among
self-employed individuals exposure to entrepreneurial risk can be associated with lower holdings of risky
assets. Less formally, some of these authors have suggested that habits and procrastination can also be part
of the story.
The research presented here is further connected with a recent literature, and even a popular debate, that
tries to understand the labor supply responses to the up and downs of the financial markets. Coronado and
Perozek (2001) and Hurd and Reti (2001) analyze this connection, but due to lack of data on the downturn the
evidence is mixed. My results can be interpreted as suggesting that if flexibility affects portfolio decisions it
should not be surprising to observe some portion of the population exercising that flexibility if the (negative)
results of the financial markets forced them to.
My research complements all this literature building upon the strong theoretical predictions of structural
models that endogenize the investment decision, the consumption decision, and the labor supply decision
of agents. Whether labor supply flexibility has an effect on portfolio choice is an empirical question, and a
very important one. This is especially true among populations, like the one I am analyzing, that hold a large
7 Svensson (1988), and Svensson and Werner (1993) solve the portfolio problem in a continuous time model with exogenous
stochastic income from non-traded assets, for example wages. Davis and Willen (2000) provide empirical estimates of the correlation between labor income and investment innovations.
3
amount of the economy’s assets, and among which labor supply flexibility is an important consideration in
all labor force related decisions.
In order to study this issue I use the HRS. This panel data set of older Americans provides good measures
of individuals’ portfolio allocations, as well as excellent direct measures of labor supply flexibility, such
as indicators of the possibility of increasing or decreasing hours of work in the current job, or whether
individuals already hold second jobs. 8
A potential problem in using the HRS is that this data set was devised to analyze the behavior of older
individuals as they move into retirement. Although the predictions from the theoretical models are independent of age, Bodie, Merton, and Samuelson (1992) emphasize the role of age in their theoretical model,
stating that the relationship between labor supply flexibility and portfolio allocation is likely to be much
weaker among older individuals. In that spirit, our results can be considered lower bounds for the influence
of flexibility measures on the holdings of risky assets by individuals. 9
The main result of the paper shows that individuals who have access to more flexibility in their jobs are
predicted to hold around 14% more risky assets than otherwise identical individuals, and those that have used
up their flexibility are predicted to hold 23% less wealth in stocks, confirming the qualitative predictions of
the dynamic models of consumption and leisure choices. These findings are obtained in a model that controls
for the selection bias caused by the asset ownership decision, and for unobserved heterogeneity using the
panel structure of the survey. Selection is a key component of the analysis. I contribute to the literature
by being one of the first empirical studies to show the advantages of using Two-Step panel data selection
models.
In the next section I present the data used for the empirical tests, and show some cross-tabulations of
the main variables and samples of interest. Section 3 presents the econometric models I use to analyze the
empirical relationship between holdings of risky assets and labor supply flexibility. Section 4 provides the
empirical results, and section 5 concludes.
8
Hurd (2002) studies the portfolio holdings of the very old, using the AHEAD data set. He focuses on asset ownership, and
transitions between ownership states.
9 It is also true, as the authors also acknowledge, that if wages become less risky over time (something very reasonable) it is
perfectly possible that we would observe increasing risk taking over the life cycle. In fact, Poterba and Samwick (1997) find, using
the Survey of Consumer Finances, that older households are more likely to hold corporate stock. Ameriks and Zeldes (2001) using
the same data set, and also administrative panel data from TIAA-CREF find that, if anything, age has a positive effect on stock
holdings. However, both sets of authors emphasize the differences across cohorts, which I cannot take into account with the data I
am using.
4
2 Data and Summary Statistics
The data I use in this research is from the Health and Retirement Study, a nationally representative longitudinal survey of 7,700 households headed by an individual aged 51 to 61 as of the first round of interviews
in 1992-93. The primary purpose of the HRS is to study the labor force transitions between work and retirement with particular emphasis on sources of retirement income and health care needs. It is a survey
conducted by the Survey Research Center (SRC) at the University of Michigan and funded by the National
Institute on Aging.10 Up to now data from the first five waves of the survey are available. The last four
waves of the data were conducted by phone using the computer assisted technology (CATI) which allows
for much better control of the skip patterns and reduces recall errors.
Death and attrition have reduced the number of participants in the survey. This can create some selectivity issues that I will mention later on. Additional individuals, who have been included, have entered the
survey later on mainly as spouses of previous respondents. The data for the respondents is being merged
from wave 5 backward to waves 4, 3, 2 and 1, and I construct a set of consistent variables on different
sources of income, financial and non-financial wealth, health, socio-economic characteristics, labor supply,
and labor supply flexibility, that will be assigned to each decision maker appropriately. In the empirical
work I use all available waves of the HRS.
Bodie and Samuelson (1989), and Bodie, Merton, and Samuelson (1992) conjecture that a possible
good measure of flexibility could be the individual’s occupation. However, using occupation as a regressor to
explain asset holdings is likely to be subject to severe endogeneity problems, given that individuals optimally
choose their occupation, and those more likely to choose occupations that offer more flexibility are likely to
make portfolio decisions very differently, leading to a spurious correlation between occupations and risky
asset holdings. They also conjecture that marital status could be a good measure, since the holdings at the
household level can be affected by the flexibility given by the possibility of having two income earners in
the household. I will test for the use of the latter variable, and instead of using occupational measures, I
choose to use more direct questions regarding flexibility, which are less likely to be the object of endogeneity
problems.
Direct labor supply flexibility questions were only asked to those working for someone else at the time
of the interview. This, however, might not represent a big weakness of our analysis, after all self-employed
individuals have, at least potentially, access to a lot of labor supply flexibility. I also believe that those not
working are probably not the best sample to use if we are to test for an effect regarding flexibility on the job.
In any case, one of the sub-samples that I will use to test the hypothesis of a positive effect of labor supply
10
See Juster and Suzman (1995), also Gustman, Mitchell and Steinmeier (1994, and 1995) or the HRS web page.
5
flexibility on the holdings of risky assets includes those self-employed, and controls for them. This strategy
is obviously not free of potentially important endogeneity problems, and I will consider it a weak test of
the hypothesis of interest. Even among those employed by someone else, I will construct a sub-sample of
respondents that consider themselves as the financially knowledgeable person of the interviewed household.
I proceed in such a way because the reported amounts of financial (risky) assets are not individual specific but
for the household, therefore having both members of a given household in the sample can create problems
in the estimation of the models. Also, it is necessary to control for the household’s position on the wealth
distribution, given its strong correlation with the holdings of risky assets, as shown, for example, in Tracy
and Schneider (2001).
It is important at this point to be more precise in describing the main variables of interest in our analysis.
The labor supply flexibility variables are dummies that take the value one if the respondent can change the
amount of hours he or she works in the regular work schedule, and zero otherwise. The definition that I
choose to focus on takes the value one if the respondent reported that he or she can increase the number of
hours in the regular work schedule, and zero otherwise. 11 I also create several other dummies that take the
value one if the person can decrease hours of work, or if the person can only decrease, or increase the hours
of work.
Another potentially interesting variable regards multiple job holding. The ideal variable would identify
those individuals that can potentially hold a second job without losing their current job. However, the only
available variable reports whether someone was holding a second job at that time. But holding a second job
at the present time can be considered a measure of lack of flexibility; if you already have a second job you
are committed to a very high number of working hours, and it is relatively difficult to increase or decrease
them. More importantly, those that choose to work a second job are probably a very different population
that are likely to invest very differently (probably much less) in the financial market.
Regarding the variables representing financial assets, there are five potentially useful variables. Individuals report the amount of money they have invested in stocks, bonds (corporate, municipal, foreign),
treasury bills and U.S. Bonds, real estate, individual retirement accounts, and other saving instruments.
Stocks represent the market value of all stock in publicly held corporations, mutual funds, or investment
trusts, not held in retirement accounts. Bonds represent the market value of the sum of all those types of
bonds. Treasury bills include T-bills, certificates of deposit, and government savings bonds. Real estate
include land, rental real estate, a partnership, or money owed to the respondent on a land contract or mortgage, and exclude the main home and second home of the respondent. Individuals’ retirement accounts are
11 Given the evidence from Coronado and Perozek (2001), and Hurd and Reti (2001) that labor supply did not strongly respond
to the upturn, meaning that they did not use the flexibility to reduce hours of work, this seems like an appropriate choice.
6
recorded at their current values. Finally, other savings are a default category for any other investments not
recorded in the previous categories. I construct three variables out of these different categories. The first
one only includes stocks. The second adds bonds, and the third adds real estate to those two. I choose not
to focus on other savings and individual retirement accounts, due to the absence of consistent information
across waves regarding the type and breakdown of the value of assets held in those investments. 12
One of the most serious drawbacks of these calculations is that we cannot assess how risky these different
financial assets are. Within each category assets with different risk properties are added together, and then
in some cases I total categories that again can have very different risk characteristics. This is not a drawback
specific to the HRS, just about any survey that tries to measure wealth and especially financial wealth, would
run into this problem. Considering all this, I have decided to consider our right hand side variable(s) as the
sum of these financial assets, and try to explain the variation in the holdings of these assets.
Summary Statistics
Using all available waves of the HRS, and exploratory descriptive statistics, we can observe in Table 1,
panels A and B, that self-employed individuals represent a very different sample of individuals from the full
sample of workers. They are more likely to be married, males, and more educated. They have more income,
a lot more net wealth, non-risky wealth (which here and in the estimations includes only vehicles, checking
accounts and savings accounts), and much more invested in risky assets. One interesting question that arises
at this time is how I can control among this population for whether the individual is more or less invested
in his or her business. One possibility is to use the response to the question as to whether the person pays
himself or herself a salary. One conjecture would be that those that pay themselves salaries are less invested
in their businesses (maybe because they have joint partnerships), and therefore invest more on risky assets.
Most self-employed respondents do not pay themselves a salary.
If in Table 1 we concentrate on those that answered the labor supply flexibility questions, that is, those
that work for someone else, around a third of them have flexibility to increase their hours of work in their
current job, and essentially the same proportion can decrease their hours of work, and around 11% holds a
second job. More than half of them can be considered hourly workers, and more than 80% of them have
health insurance through their employers.
From Tables 1 and 2, comparing stock owners with those that do not own stocks (either among all employed individuals or only among employees) reveals marked differences along just about every dimension.
Stock owners are more likely to be white, male, married, and more educated. They also have higher earnings, non-risky wealth, are less likely to be hourly workers, have more flexible labor supply, and self-report
themselves to be in better health.
12
Ameriks and Zeldes (2001) analyze portfolio holdings over the life-cycle using administrative data from retirement accounts.
7
The sub-population that holds second jobs at the time of the interview has comparatively lower levels of
risky assets. This seems to support the conjecture that holding a second job can be considered a measure of
lack of flexibility, and probably correlated with unobservables (financial problems that force the individual
to search for a second job) that if anything reduce the likelihood of holding risky assets. This last observation
is supported by the fact that this sub-group is substantially more educated than any other group (except for
stock-owners), a characteristic, that other things equal, I would expect to have a positive effect on holdings
of risky assets.
Finally, in Table 2 the sub-population that can increase their hours of work holds substantially more
risky assets. These individuals are more likely to be white, male, and married. They worked a bit less in
the year before the interview, but earned more than the full sample of financially knowledgeable employees.
They also self-report themselves to be in better health.
The conclusions from this exploratory analysis is that stock-owners are a very different sample, and
the ownership decision, which selects our sample, deserves separate analysis. There is some evidence of a
positive effect of labor supply flexibility, and a negative effect of holding a second job, but the unconditional
differences in observables dominate the analysis at this stage.
3 Cross-Section and Panel Data Models
Using the data presented in the previous section the empirical problem is to find whether there is a statistically significant relationship between labor supply flexibility and the asset allocation decisions of the
individuals in our sample. The dynamic models I discussed in the introduction indicate that more labor
supply flexibility should translate into more investments in risky assets.
To satisfactorily answer this question a number of econometric questions need to be tackled. First of all,
we want to appropriately control for observed heterogeneity. This is not a trivial task since omitted variable
biases can be an important problem regarding portfolio allocations. For example, an important issue in this
research is to control for the position of individuals in the wealth distribution, but we want to accomplish
this without including as independent variables (possibly) endogenous measures of the dependent variable.
Second, and probably more important, sample selection issues are potentially very problematic. I am
interested both in understanding financial the ownership and the distribution of financial assets. But we have
to take into account that to estimate the latter we have to do it conditional on ownership and this can bias the
coefficients as if we were incurring a specification error (Heckman 1979). It is not very difficult to control
for this selection problem in cross-sections, but it becomes a more involved problem in panel data models,
as is the case in my sample. In fact few empirical applications tackle this issue appropriately.
Third, we want to take into account the unobserved heterogeneity potentially present in our characteriza8
tion of the econometric model. If we do not control for the unobserved components we will be confounding
partial and total effects of our variables of interest. Panel data sets allows us to model explicitly how those
unobserved components enter the econometric specification, and we can choose to include them as a fix
effect or as a random variable, and test the different specifications. Notice, however, that the panel data
structure complicates the sample selection and the attrition problems. Appropriately controlling for these
biases in the panel models presents the researcher with interesting methodological challenges. I will not
directly account for attrition biases in the estimations, but the use of a panel data selection correction model
will ameliorate this problem.
Once we have decided which variable to use as the dependent variable (I will use total net investments
in stocks), and which indicator of flexibility we want to use as the main independent variable of interest
(I will use an indicator of whether individuals can increase their hours of work in their jobs), we can start
by estimating in a given cross-section or the pooled sample of observations (controlling for clustering, see
Deaton 1997) the following model
ln Ai
α1
AOi
α2
Xi β
u1i
(1)
and
Zi γ
u2i
(2)
where in (1) the set of individual characteristics Xi consists of various socio-economic and demographic
variables, and other variables I will describe below, one or more labor supply flexibility indicators, and A i
represents a measure of the risky asset(s) currently held by the individual. In (2) AO i is an indicator of
whether the individual owns a risky asset(s), and u 1i and u2i are not independent, and where Zi can be equal
to Xi but will be larger in the empirical application, which will help in identifying the model. In this setting
the selection rule is potentially not independent of the behavioral function being estimated. However, I
argue, following Blundell and Meghir (1987, p. 180-181), and Maddala (1988, p. 286), that the Tobit
characterization, which is a common alternative to this set up, is not the preferred framework to estimate
this type of model since it does not seem appropriate to assume that the selection rule, the probability of
owning the asset, and the process that leads to own more or less stocks (or other financial assets) are the
same. Also, we cannot forget that the lack of observability of all the values of the dependent variable in
the main equation is the result of a choice by individuals, not the result of censoring. The empirical results
will make clear that this more general structure of the problem provides for a better understanding of the
underlying decisions of individuals.
It is fairly straightforward to estimate the full model by Maximum Likelihood or by standard Two-Step
procedures. It is important to emphasize that the study of financial asset ownership is important on its own,
and the HRS allows us to perform both a cross-section and a panel data analysis of this issue to complement
9
the study of the levels of the portfolios of risky assets in the HRS sample.
The HRS provides us with repeated observations of the same individuals. This allows us to control for
potential unobserved components that could enter our econometric model. Our main equation of interest
can now be written as
ln Ait
α1
Xit β1
µi
νit
(3)
where µi represents the unobserved heterogeneity component, and the ν it are the idiosyncratic disturbances.
This model can be estimated assuming either no correlation between observed explanatory variables and the
unobserved effect (random effects), or allowing for arbitrary correlation between the unobserved effect and
the observed explanatory variables (fixed effects). We can then test whether the random effects specification
or the fixed effect specification is more appropriate, and whether the former is more appropriate than the
pooled OLS regression.13
We have, therefore, an unbalanced panel of respondents, and individuals are selected into the sample,
and non-randomly drop from the panel of individuals that own financial assets. One possibility to take this
selection problem into account is to simply extend the classic Heckman (1979) sample selection correction
approach where the first stage is a probit specification with no individual component and the second stage
accounts for the individual component. This is suggested by Hsiao (1986), and it is fairly straightforward,
but misses the interesting (and potentially important) issue of selection into the panel. Also, it is tough to
justify why there is an individual component in the equation of interest but not in the selection equation.
It is, therefore, natural to think that we can allow both the equation of interest and the selection equation
to have individual specific components, on top of the usual error component, which will allow us to take
into account unobserved heterogeneity. This will result in obtaining two correction terms, which I will be
able to trace back to different reasons for the selection into the sample. 14
In this case I am especially interested in estimating the effect of a number of time invariant regressors
both on the selection equation and the equation of interest, so it can be natural to model the individual
components as random effects. Also, following the suggestions of Jensen, Rosholm, and Verner (2001),
and the work of Mundlak (1978) and Zabel (1992), it is possible to account for some correlation structure
between exogenous variables and unobserved components within the random effects model. On the other
hand, Jensen, Rosholm, and Verner (2001) indicate that estimating a fixed effect of the main equation after
13
See Wooldridge (2002) for an up to date and illuminating presentation of these issues.
There are surprisingly few empirical papers using these techniques, with the exception of the examples in Vella (1998), and
the work of D’Addio, De Greef, and Rosholm (2002) who study unemployment traps in Belgium, but choose to estimate the model
by Maximum Likelihood, which, although fairly well behaved in Monte Carlo experiments, is comparatively less robust than TwoStep estimators and imposes additional distributional assumptions. A recent edition of an econometric software package, Limdep
8.0, has some code that in principle can estimate this model by Maximum Likelihood. However, after experimenting with it I had
problems of convergence, which force me to simplify the correlation structure of the problem, and even under those simplifications
some of the problems seem to persist.
14
10
the correction can result in better behavior in terms of consistency and efficiency even in the presence of
correlation between the exogenous variables and the effects, and with both logistic and normal errors. 15
The presentation of the model below follows mainly Vella and Verbeek (1999), Jensen, Rosholm, and
Verner (2001), and Vella (1998), and in part Verbeek and Nijman (1992), and Nijman and Verbeek (1992,
1996). Miniaci and Weber (2002) also provide a nice characterization of some of these issues in portfolio
analysis.
Consider the equation of interest (3) where the usual error component structure is considered where
iid 0 σ2µ and νit
µi
IIN 0 σ2ν The selection equation can be written
α2
AOit
Zit γ
εi
ηit
(4)
where εi and ηi t are independent of each other and also of Xit , and we have
εi
IIN 0 σ2ε and ηit
IIN 0 σ2η This structures allows us to consider the case where the individual effects in both equations are correlated, and that is also true of the two error terms. So I can write
cov εi µi σεµ
cov ηit νit σην
In a Two-Step procedure I will be estimating these covariances once I construct the appropriate correction
terms. The selection correction terms correspond to:
¯i E µi AO
σεµ A1
¯i E νit AO
σην A2 It is possible to write separate expressions for A 1 and A2 , but I can also follow Vella and Verbeek (1999) to
write from (3)
αit
µi
υit
εi
νit
and from (4)
ηit
15 See Dustmann and Rochina-Barrachina (2000) for an empirical comparison of models where the selection equation is assumed
to have fixed effects. See also Kyriazidou (1997), and Kyriazidou (2001). Jensen, Rosholm, and Verner (2001) find that the latter
approaches behave less robustly than the one followed in this paper.
11
and υi is the T vector of υit ’s
υi Xi
NID 0 σ2ε υ
E αit Xi υi τ1 υit
υ¯i
T
T
∑ υit
1
σ2η I τ2 υ¯i
(5)
individual specific average t 1
(5) has the form of the former correction terms, where τ 1 and τ2 would be the additional parameters to be
estimated. Notice that (5) is an assumption that holds for normal distributions.
Now, I want to compute υit .
E εi
υit
¯i ηit AO
εi
¯i εi f εi AO
¯i dεi
E ηit AO
(6)
¯i εi is the cross-sectional generalized residual (Gourieroux et al. 1987)
where E ηit AO
φ Zit γ̂ εi AOit
Φ Zit γ̂ εi ¯i εi E ηit AO
φ Zit γ̂ εi 1 AOit 1 Φ Zit γ̂ εi (7)
And
f AOit εi Zit f AOit Zit ¯i f εi AO
(8)
is the density of εi given selection and where the denominator is the likelihood contribution in the Random
Effects Probit for an individual:
f AOit Zit f AOit Zit εi f εi dεi
and
f AOit εi Zit f AOit Zit εi f εi here the likelihood contribution in the cross-sectional case is:
f AOit Zit εi ΠtT 1 f AOit Zit εi Once I obtain υit , I compute the individual specific average υ¯i and estimate
ln Ait
α1
Xit β1
12
υit θ1
υ¯i θ2
ui
(9)
by OLS or Random Effects, depending on whether we think u i , an additional individual component which
is uncorrelated with the individual component in the selection equation, should be added. 16
Notice that estimating (7) is fairly straightforward, and then to estimate (8) I only need to integrate out
the distribution of the unobserved component and can follow the arguments in Hewett and Montgomery
(2003) that present a related problem as an application of Bayes Rule. After that, only a one dimensional
integral is needed to calculate (6) to obtain the correction terms. 17
Finally, as I mentioned above, and as Jensen, Rosholm, and Verner (2001) emphasize, including in (4)
the individual specific averages over time of the exogenous variables as additional regressors we provide
a correction for the possible correlation between the exogenous variables and the individual effects. There
is no guarantee that this will work in a non-linear estimation problem, but the Monte Carlo results of the
authors above suggest that it can improve the consistency and the efficiency of the Two-Step estimator.
4 Empirical Results
In this section I present the results of several types of models. First, I estimate equation (1) by OLS using
the pooled sample of individuals that own stocks. Second, I use Heckman (1979) Two-Step selectivity
correction technique to incorporate equation (2) into the estimation. Third, I estimate panel data sample
selection models, (3) and (4), and (9), allowing for unobserved heterogeneity.
I estimate these models for two different sub-samples. The first sub-sample includes both self-employed
and employed individuals, the second sample only includes those working for someone else. The latter
sample is the one that incorporates the flexibility indicator and where I ultimately test whether labor supply
flexibility, as I define it, matters. As part of the process I also estimate for both sub-samples cross-section
and panel data binary choice models of the ownership indicator.
A non-trivial task in the empirical work has been to decide the right combination of observed controls.
I have followed the literature that analyzes risky assets ownership and levels, but I encountered several
problematic decisions. For example, deciding how to control for socio-economic status is complex. Income
and wealth are the obvious candidates, but the latter has to be handled with extreme care since we do not
want to include proxies for our dependent variable among our regressors. This is why I construct a series
of indicators of wealth that do not include any assets that could be considered risky. Also, on top of the
traditional demographic controls, I include self-reported health measures, which seem to have a significant
16
Notice that if there is no individual component in the selection equation then (6) reduces to the Inverse Mills’ ratio in Heckman
(1979), assuming, as I have done above, that the individual component of the main equation is either absent or uncorrelated with
the idiosyncratic component of the selection equation.
17 I have estimated the Random Effects Probit model and constructed the correction terms using Fortran 90. Special thanks to
Mark Montgomery for his help with the code.
13
effect on ownership but little on the level of asset holdings. 18 I also include a variable that takes the value one
if the person plans to work full-time by age 62. This variable has been shown to have considerable predictive
power in other contexts, and as we will see, also plays a significant role in our estimations. Finally, I also
include an indicator of whether the individual is an hourly worker.
4.1 Self-employed and employees sub-sample
Table 3 provides pooled cross-sectional and panel regression estimates of (1) and (3), where the dependent
variable is the logarithm of the value of stocks in the portfolio, using the sub-sample of employed and selfemployed individuals that are also the financially knowledgeable persons in their households. As I explained
in Section 2, flexibility measures are not available for all these individuals, and in this case the main proxies
for flexibility are dummies that take the value 1 if the person is self-employed and receiving a wage or
self-employed and receiving profits, and zero otherwise. The hypothesis is that self-employed individuals
receiving a wage would invest more in risky assets since they have flexibility but are likely to be less invested
in their businesses. Those that pay themselves a profit are likely to be more invested in the business, and in
line with the results of Heaton and Lucas (2000b), we would expect a lower exposure to risky assets.
The first column shows the estimation of a standard pooled OLS regression, the second one a fixed effect
panel estimation, and the last column provides a random effects panel estimation. The specification tests
indicate that the random effects specification is preferred to the OLS, thus the classical one-constant model
is soundly rejected. Also the Hausman test rejects the Random Effects in favor of the Fixed effects, however,
the fixed effects estimates are noisy, and following the intuition in Wooldridge (2002) I consider the Random
Effects our preferred specification. Notice that the identification of the coefficients of the self-employment
indicator comes from the variation over time of this coefficient for each cross-sectional unit (individual).
This is not very common in this sample, therefore it is not surprising that the Fixed Effects performs poorly.
Also the Random Effects specification allows us to estimate the effects of interesting non time-varying characteristics. In the Random Effects specification the estimates indicate that self-employed individuals who
receive wages hold higher levels of risky assets after controlling for observed and unobserved heterogeneity, and those that receive profits accumulate less risky assets. However, in all cases the indicators are not
significant, suggesting that compared with employed individuals these effects are not strong enough. More
encouraging results come from including an indicator of whether the individual already holds a second job.
As mentioned earlier this is a proxy for lack of flexibility. 19 In this case the effect is significant, and negative,
18
Rosen and Wu (2003) focus on the effect of health status on portfolio decisions. Our results are in line with theirs regarding
participation in the market of risky assets, but rather different regarding total accumulation. The differences come from the fact that
they estimate a pooled Tobit model without a separate characterization of the process of ownership and level of investment.
19 Holding a second job can also proxy, for example, for limited future earnings power, or lower wage growth expectations,
interpretations that are consistent with the results reported throughout the paper.
14
confirming the hypothesis. We will see later that this is a clear result in all the specifications I report in this
paper.
Being white, more educated, and married have positive effects on the amount of risky assets the household owns. Where the married dummy could again be proxying for some labor supply flexibility if we take
a household perspective (See Bodie, Merton, and Samuelson 1992), however, this latter indicator is not
significant. Individuals in better health are likely to hold more financial assets, interestingly this result will
be reversed once we take into account selection, but in all cases the level of significance is very marginal.
Another interesting result that seems to be general to all the specifications I have tried, in this and other
sub-samples, is that a variable that measures the self-reported probability of working full-time past age 62 is
strongly negative correlated with the holdings of risky assets, other things equal. One possible explanation
for this result could be that those individuals who expect to work full time past age 62, the Early Retirement
Age, are probably those that will need to, due to some bad shock to their wealth or to past income and
employment. They simply cannot afford to retire early. 20
Of course all these results are exposed to selection bias. Therefore, I need to estimate the selection
equation and make the appropriate econometric corrections. Table 4 provides estimates of a pooled probit
and a Random Effects panel data Probit of the stock ownership indicator. 21 These results show that on top
of the variables we used in Table 3, the effect of smoking, which is negative, and moderate drinking, which
is positive, and non-risky wealth, also positive, are very significant. Furthermore, individuals without health
insurance are much less likely to hold stocks and the same is true of hourly workers. Those in better health
are more likely to hold stocks and, rather surprisingly, time indicators for the beginning of the 1990s indicate
a higher probability of owning stocks. This last result is quite robust across samples and specifications. My
interpretation of this is that although the level of stocks clearly grew strongly during the 1990s (due mainly
to the impressive returns of the bull market), other things equal, the same is not necessarily true of ownership
among this population. In fact if I include dummies for waves 3 to 5 they are estimated to have negative and
significant coefficients. These results are in line with the aggregate results of Tracy, Schneider, and Chan
(1999), and Tracy and Schneider (2001).
This first test of the relationship between labor supply flexibility and portfolio choice is, however, mixed.
Self-employment is a measure of flexibility but it can also be a measure of many other things that are
potentially correlated with financial asset holdings. It seems that endogeneity problems are very difficult
to avoid given that is virtually impossible to find an appropriate instrument for the self-employment status.
This is true even after controlling for unobserved heterogeneity. However, the second job indicator provides
20
See Benı́tez-Silva and Dwyer (2003) for a study of retirement expectations formation using the same data set.
These estimates include as covariates the same variables as in Table 3 plus the variables labeled 2 to 6. Therefore, the
correction term(s) will be identified on more than just functional form.
21
15
a result in favor of the hypothesis that flexibility matters.
The next step is to take into account the selection process to estimate (3). Table 5 provides four different specifications of the selection corrected results. The first two columns correct the selection bias by
estimating (2) in the first stage, and then in the second stage estimate either a OLS or a Random Effects
specification. The last two columns estimate (4) in the first stage, and therefore require the inclusion of two
additional terms in the estimation of the equation of interest, which can then be estimated by OLS or by
Random Effects, depending on whether we think there is an additional individual component uncorrelated
with the individual component of the selection equation. 22
In all cases the selection correction terms are extremely significant suggesting that the unmeasured
factors related to the probability of owning stocks are correlated with the unmeasured factors that affect
the level of stocks owned by individuals. Even more interesting, it is rather surprising to see in the first
two columns of the table that the Inverse Mills’ Ratio is actually negative. This means that the unmeasured
factors that make an individual more likely to own stocks make them more likely to own less stocks. The
reason for this is the skewed distribution of stock holdings, with most individuals owning a relatively small
portfolio, and with some others owning literally millions in the stock market.
In clarifying this further it is important to implement the Two-Step estimator for panel data selection that
breaks down the selection into two components. Columns three and four show the estimates of the two separate selection terms, υit and υi . Selection into the market for stocks has an individual specific component,
υi (which reflects stock ownership differences across individuals) which is positive, but an idiosyncratic
component, υit (which reflects the stock ownership differences over time for a given individual) which is
negative. This means that if we compare those that own with those that do not own across individuals, the
unmeasured factors that make you more likely to own have a positive effect on the level of holdings, but
if we compare ownership and non-ownership for a given individual over time the factors that increase the
probability of participation actually make you likely to own less stocks. In essence these two terms allow
us to separate the effect of being an “opportunistic stock market participant”, and the effect of being a “true
stock holder”. This is the first study to obtain this surprisingly intuitive result regarding participation in the
stock market in the 1990s. We will see below that when we concentrate on employees this powerful result
is also present.
Looking at the effects of the other variables selection also matters. The only demographic variable
that is still significant is having a professional degree, which increases the stock holdings significantly. On
the other hand having more non-risky wealth still affects stock holdings in a significant and positive way,
22 The specifications in columns 2 and 4 are preferred to the ones in columns 1 and 3, respectively, given the results of BreuschPagan Lagrange Multiplier tests (not shown). I also estimated the second stage by fixed effects following the suggestions of Jensen,
Rosholm, and Verner (2001), the main results were essentially unchanged.
16
and surprisingly the same is true for those that do not have health insurance. This latter result reverses
the findings of the uncorrected regressions, suggesting that once I control for the negative effect that lack
of insurance has on participation in the market, those “risk lovers” in terms of health insurance are also
more invested in the stock market. Considering the proxies for labor supply flexibility, there is still a mixed
picture. The self-employed indicators are not significant, although have the right signs, but the second job
indicators is still significant and related to lower holdings of stocks. Finally, notice that after I control for
sample selection into the panel and unobserved heterogeneity, health variables are no longer significant, and
in some cases have the opposite sign compared with the uncorrected results. This suggests that health effects
come mainly through the participation decision.
Table A.1. in the appendix provides some sensitivity analysis for the previous results. All the results
provided are the product of estimating a Random Effects Probit model of the selection, and then estimating
a random effects of the equation of interest, as in equation (9). In specification 1 I substitute the two selfemployment indicators for a single indicator, the results are essentially unchanged. Specification 2 changes
one of the more significant variables, the probability of working past age 62, to the same type of probability
to working past age 65. The number of observations decreases, but the results do not change. Finally,
specifications 3 and 4 change the dependent variable, with specification 3 adding bonds and real estate to
stocks, and specification 4 adding only bonds. The main difference appears in specification 3 where the
self-employment indicators are highly positive and significant. This is due to the large accumulation of
investments in real estate by self-employed individuals compared with employees.
4.2 Employees sub-sample
The HRS provides better measures of labor supply flexibility than the ones I have used for the full sample of
workers above, but the price we have to pay is the dropping of self-employed individuals from the sample.
Given how different they are from respondents that work for someone else, and our objectives, this might
not be an important drawback. In the tables that follow we concentrate on employees that work for someone
else and who are also the financially knowledgeable persons in their households.
In Table 6 I use as the dependent variable the sum of stock holdings, and an indicator of labor supply
flexibility a dummy that takes the value one if the respondent self-reports that they can increase the hours
of work in their current job. I also include the indicator for holding a second job. In the first set of columns,
I run the OLS regression without the selectivity correction, the second set shows the fixed-effects estimates
and the third set of columns the random effects estimates. In the pooled uncorrected OLS the effect of
the flexibility variable is large and significant, predicting that those with flexible labor supply would hold
almost 17% more wealth in stocks, but those that already hold a second job would hold around 34% less
17
stocks. The specification test rejects the OLS model in favor of the Random Effect specification, and the
Hausman test rejects the hypothesis that the regressors and the unobserved components are uncorrelated.
However, as in the previous tables, the Fixed Effect estimates are quite noisy. The flexibility indicator does
not change much but it is more imprecisely estimated, something not surprising considering that this is not
a variable that changes a lot overtime for a given individual. All this, and the ability to estimate the effect
of time-invariant variables, convinces me into choosing the Random Effects as our preferred specification.
In this specification the effect of Labor Supply flexibility on the amount of stocks held by the household is
again almost 17%.
All these estimates are affected by possible selection bias. I therefore proceed to estimate the selection
equation and then make the appropriate corrections. The ownership estimations in Table 7 do not present
big surprises. Notice, however, that the flexibility indicators play a small and insignificant role in these
equations, suggesting that once we correct for selectivity in the panel estimates, we are likely to see relatively
little movement in these coefficients. Again notice the expected positive effect of having a college education
in the participation decision, and the powerful effect of socio-demographic variables, the negative effect of
being hourly paid, and the negative effect of self-reporting a higher probability of working full-time past
age 62. Also, those that work a higher proportion of months in the previous year are more likely to hold
assets. I again find here the interesting presence of time effects that indicate that participation among this
aging population was actually on its way down, other things equal, during the 1990s. This is not to be
confounded with the very positive effect that the late 1990s have on total holdings of stocks, which is clear
in the uncorrected and the corrected results.
Following the same structure as Table 5, Table 8 provides five specifications of selection corrected results. The first two specifications estimate a pooled probit model in the first stage, and then a standard
Heckman correction of the second stage using either OLS or a Random Effects model. The last three specifications estimate a Random Effects Probit model in the first stage, and then follow Vella and Verbeek (1999)
to construct the two correction terms to account for the panel sample selection. The third specification estimates an OLS model in the second stage and the fourth specification a Random Effects model in the second
stage which means that I allow for an individual component that might not be correlated with the individual
component in the selection equation. The fifth follows the recommendations in Jensen, Rosholm, and Verner
(2001), and the earlier contributions of Mundlak (1978) by including in (4) the individual specific averages
over time of the exogenous variables as additional regressors, to provide a correction for the possible correlation between the exogenous variables and the individual effects. In Monte Carlo experiments this has
been shown to improve the consistency of the estimator. In this case results do not change much, and in the
first stage simple F-tests show that not all additional regressors are significant suggesting that the Mundlak
18
correction is not necessarily appropriate. 23
Although the message from all specifications is essentially the same, I concentrate on the fourth specification, which estimates equation (9), which I believe better captures the underlying structural model of
stock ownership and stock holdings. The main reason is because it allows me to separate two different
selection effects present in the panel, and also because it is realistic to think that there is an additional
individual component in the main equation which is not correlated with the individual component of the
selection equation.24 The main result is that the flexibility indicators are sizable and statistically significant
at the 5% level. Individuals with the flexibility to increase their hours of work hold on average 14% more
wealth in stocks, while those that already have a second job hold around 23% less stocks. Notice that we
are conditioning on a large array of variables, which in some cases see their effects reversed from the ones
they had in the uncorrected estimations. Having a professional degree has a large positive and significant
effect and so does having higher non-risky wealth, and not having health insurance. As in Table 5 I believe
this latter result captures the fact that the health insurance effect works mainly through the participation
decision, and conditional on it, those that hold more stocks are individuals who are likely to behave riskily
in other markets, like the health insurance market. The probability of working full time past age 62 affects
stock holdings negatively, proxying for individuals that cannot afford to retire early maybe due to previous
bad shocks to wealth or income. Those with a health limitation are likely to hold significantly less stocks,
but other self-reported health measures are not significant. This comes to reflect the fact that most of the
action of health measures goes through the participation decision. Notice also the highly negative effect of
the time indicators for the early part of the 1990s, which again captures the impressive growth that stocks
had in the last part of the decade. The contrast with the declining participation during the decade is a rather
unexpected result.
Finally, the two selection terms are very significant, and have opposite signs, reflecting a very insightful
finding regarding the nature of the panel selection into owning stocks for our sample. Selection into the
market for stocks has a positive individual specific component, υ i in Table 8, which reflects stock ownership
differences across individuals, and a negative idiosyncratic component, υ it in Table 8, which reflects the
stock ownership differences over time for a given individual. This means that if we compare those that own
stocks with those that do not own stocks across individuals, the unmeasured factors that make those more
likely to own have a positive effect on the level of holdings, but if we compare ownership and non-ownership
for a given individual over time, the factors that increase the probability of participation actually make those
likely to own less stocks. This suggests that some unmeasured factors that influenced the decision to enter
23
The last three specifications do not adjust, for the moment, the standard errors to account for the added generated regressors.
As with the full sample of employed individuals the Breusch-Pagan Lagrange Multiplier tests always reject the specification
that estimates an OLS in the second stage in favor of one that estimates a Random Effects model.
24
19
the market in light of the performance of the financial assets in the late 1990s make an individual more
likely to own a small portfolio during the period of analysis (opportunistic buyers), while the unmeasured
factors that make an individual more likely to participate all through the panel (true market participants),
make those more likely to own a larger amount of wealth in stocks. As far as I know this is the first study to
obtain this intuitive result regarding participation in the stock market in the 1990s.
Table A.2. in the appendix provides some sensitivity analysis of the results for the sub-sample of employees, using the panel sample selection and estimating a random effects specification in the second stage.
In the first specification where I change the covariate that reflects the probability of working full time past
age 62 to the probability of working past age 65, we lose some observations, the flexibility indicators are a
bit larger, but the main results are unchanged. In the second specification I change the covariate non-risky
wealth to a broader definition that includes assets from businesses and other savings. The results are again
basically unchanged. Specification 3 in the Table provides sensitivity analysis on changing the dependent
variable to a broader definition of risky assets, that on top of stocks include bonds and real estate holdings.
This time the results are a bit less strong, but still significant in the case of the flexibility to increase hours
indicator, in favor of the importance of labor supply flexibility in the level of holdings of these assets. The
last specification changes the dependent variable to include only stocks and bonds, the results are essentially
the same as in the preferred specification in Table 8.
From all these results, I conclude that there is empirical evidence that labor supply flexibility has a
positive and significant effect on holdings of risky assets (especially stock holdings), as the work of Bodie,
Merton and Samuelson (1992) predicted. The estimation results predict an increase of around 14% in the
holdings of risky assets if flexibility to increase hours of work is present, and a decrease of 23% if that
flexibility has been used up, once I control for panel sample selection and unobserved heterogeneity. These
results provide one of the first empirical evidences of a strong connection between labor supply flexibility
and portfolio choice decisions. It seems that individuals who have more opportunities to increase their labor
supply are more likely to have more risky investments given that the labor supply flexibility provides some
insurance to their riskier capital investments. The Two-Step panel selection estimation strategy also provides
interesting results regarding the nature of the selection into stock ownership during the 1990s among this
population, and shows the usefulness of the the more sophisticated estimation technique used in this paper.
5 Conclusions
In this paper I have presented one of the first empirical tests of the relationship between labor supply flexibility and portfolio choice. I have tested the theoretical predictions of two different but related dynamic
models of labor supply and consumption/saving decisions, using the Health and Retirement Study.
20
The empirical results show that individuals that have more labor supply flexibility are likely to hold
higher levels of risky assets (around 14% more), and those that can be considered to have exhausted that
flexibility own around 23% less wealth in stocks. These results are the product of correcting for panel sample selection into asset ownership and unobserved heterogeneity following the work of Vella and Verbeek
(1999), the recommendations of Jensen, Rosholm, and Verner (2001), and the earlier work of a number of
authors. The Two-Step panel data model allows us to separate the selection effect on assets holdings of two
types of stock owners, the “opportunistic buyers”, and the “true stock owners”. The first effect reduces stock
holdings, the second increases it.
I also contribute to the growing literature that tries to understand the relatively low participation in
financial markets by American households. Part of the explanation (although still a lot remains unexplained)
is that we cannot forget that labor supply can be an insurance device, and that the decision to participate and
the decision to hold more or less wealth in stocks are connected but remain fairly different processes. I also
find that even during a period of unprecedented growth in stock values, participation in the stock market (not
including the participation through private pensions) among older Americans is not necessarily positively
affected, other things equal. This suggests that the decline in stock market values that started in the second
part of 2000 might have affected the non-pension wealth of this population less than generally believed.
The connection between labor supply and portfolio allocation has been studied less than each of these
topics separately, and until recently very few models tried to tie these two types of decisions together. Bodie
and Samuelson (1989), and Bodie, Merton and Samuelson (1992) emphasized this relationship and show
how ex-post labor supply flexibility has an effect on the ex-ante portfolio allocation of utility maximizing
individuals. In a different context, Benı́tez-Silva (2002) shows that a stochastic dynamic life cycle model
with endogenous labor supply, consumption/saving and annuity decisions, predicts that risky assets are also
preferred by utility maximizers.
The data that I use provides excellent measures of wealth, and a set of rarely available direct measures
of labor supply flexibility. Using several definitions of risky assets, and different measures of flexibility, I
am able to provide empirical backing to the dynamic theories. An extension outside the scope of this paper,
and one that would require the use of an alternative data set, would be to check how robust these results are
to the use of a data set that was devised to focus on older individuals. Bodie, Merton, and Samuelson (1992)
conjecture that their predictions should hold more strongly among young individuals, given that they have
more opportunities to insure against adverse portfolio outcomes through their future labor supply. However,
the lack of appropriate measures of labor supply flexibility and good measures of financial wealth make this
extension of the project a challenging one.
The fact that I have been able to find a relationship between labor supply flexibility and portfolio choice
21
makes even more interesting the project of exploring how this translates at the macroeconomic level. Is it
the case (as Bodie and Samuelson 1989 conjecture and some recent research and debate suggests) that we
would expect to see stabilizing labor market responses to shocks in securities markets? If so we need to
understand better the connection between labor supply and asset allocation decisions.
Also, it might be interesting to explore more in depth the relationship between human capital formation
and labor supply flexibility. Flexibility is potentially valuable for individuals, given that it allows them to
have some insurance against bad investment outcomes. Additional flexibility can be acquired through the
career opportunities that more education and training give to individuals. However, very few models of
human capital investment have tried to incorporate this connection.
Finally, we can not ignore how important labor supply flexibility is for older individuals. Some recent
evidence seems to indicate that older workers are likely to extend their labor force participation beyond
traditional retirement ages. However, flexibility emerges as a decisive characteristic which older workers
take into account when making labor supply decisions. Interestingly, this flexibility will not only have an
effect on labor supply, but also on wealth accumulation and de-accumulation decisions of individuals that
hold most of the financial assets in the economy. 25 The study of the links between labor economics and
finance (both at the micro and macro level) is likely to be one of the fastest growing research topics that both
fields will be tackling in the next years.
25
See Davies (1981), Bernheim (1987), Alessie, Lusardi, and Kapteyn (1999), Banks et al. (1998), and Moore and Mitchell
(2000) for a discussion of savings around the time of retirement.
22
Table 1. Panel A: Means and Standard Deviations. Employed and Self-Employed
Respondents.
Variable
# of Observations
Age
White
Male
Married
Bachelor Degree
Professional Degree
Respondent Income
(in $1,000 of 1992)
Net Worth
(in $ 1,000)
Non-Risky NW
(in $ 1,000)
Stocks
(in $ 1,000)
Bonds
(in $ 1,000)
Real Estate
(in $ 1,000)
Stock Ownership
Hourly Worker
Self-Empl. with Wages
Self-Empl. with Profits
% Months Worked
Flexible to Increase Hwk.
Flexible to Decrease Hwk.
Second Job
Employee
Sub-Sample
15,700
Self-Employed
Sub-Sample
3,795
Stock Owners
57.92
5.08
0.71
0.45
0.50
0.50
0.66
0.47
0.26
0.44
0.10
0.30
28.72
32.35
165.51
319.56
33.65
123.74
75.29
260.92
5.38
31.28
93.18
216.20
0.26
0.44
0.56
0.50
0.00
0.00
0.00
0.00
0.94
0.20
0.33
0.47
0.32
0.47
0.11
0.31
59.16
5.38
0.82
0.38
0.66
0.47
0.77
0.42
0.33
0.47
0.13
0.34
49.87
134.74
521.94
1159.89
166.42
440.80
168.31
839.43
25.72
153.04
289.61
515.13
0.35
0.48
0.00
0.00
0.36
0.48
0.79
0.41
0.94
0.20
—
0.00
—
0.00
0.12
0.32
23
5,372
Don’t Own
Stocks
14,123
Second-Job
Sub-Sample
2,167
58.20
5.12
0.90
0.30
0.64
0.48
0.78
0.42
0.45
0.50
0.19
0.39
48.46
97.02
485.64
996.39
118.13
353.97
98.54
478.27
28.86
136.65
200.42
430.50
1.00
0.00
0.30
0.46
0.09
0.28
0.20
0.40
0.96
0.17
0.35
0.48
0.34
0.47
0.12
0.33
58.15
5.18
0.67
0.47
0.49
0.50
0.65
0.48
0.21
0.41
0.08
0.26
26.26
45.13
135.88
286.56
37.19
153.55
—
0.00
1.76
17.62
118.50
267.22
0.00
0.00
0.50
0.50
0.06
0.24
0.14
0.34
0.94
0.21
0.33
0.47
0.31
0.46
0.11
0.31
57.10
5.03
0.76
0.42
0.58
0.49
0.69
0.46
0.38
0.49
0.17
0.37
36.51
61.97
286.68
635.33
82.61
323.79
71.52
228.77
16.88
102.45
192.45
429.38
0.31
0.46
0.41
0.49
0.07
0.26
0.16
0.37
0.95
0.18
0.33
0.47
0.35
0.48
1.00
0.00
Table 1. Panel B: Means and Standard Deviations. Employed and Self-Employed
Respondents.
Variable
# of Observations
Employer Health Insurance
Private Health Insurance
No Health Insurance
Have thought about retirement
Prob. Working after 62
Prob. Living to 75
Health Limitation
Excellent Health
Very Good Health
Good Health
Smoker
Drinker
wave1
wave2
wave3
Employee
Sub-Sample
15,700
Self-Employed
Sub-Sample
3,795
Stock Owners
0.81
0.39
0.17
0.37
0.08
0.27
0.48
0.50
0.48
0.40
0.67
0.28
0.18
0.39
0.21
0.40
0.34
0.47
0.31
0.46
0.23
0.42
0.58
0.49
0.27
0.44
0.22
0.41
0.19
0.40
0.35
0.48
0.34
0.47
0.17
0.37
0.41
0.49
0.59
0.39
0.70
0.26
0.23
0.42
0.26
0.44
0.34
0.47
0.27
0.44
0.19
0.40
0.64
0.48
0.25
0.43
0.22
0.41
0.20
0.40
24
5,372
Don’t Own
Stocks
14,123
Second-Job
Sub-Sample
2,167
0.77
0.42
0.23
0.42
0.03
0.16
0.51
0.50
0.48
0.40
0.70
0.25
0.18
0.39
0.28
0.45
0.40
0.49
0.24
0.43
0.17
0.37
0.74
0.44
0.28
0.45
0.27
0.44
0.17
0.38
0.76
0.43
0.19
0.39
0.12
0.33
0.45
0.50
0.50
0.40
0.66
0.29
0.20
0.40
0.19
0.39
0.32
0.47
0.33
0.47
0.24
0.43
0.54
0.50
0.26
0.44
0.20
0.40
0.20
0.40
0.77
0.42
0.21
0.41
0.08
0.28
0.48
0.50
0.56
0.39
0.69
0.26
0.20
0.40
0.27
0.44
0.38
0.48
0.27
0.44
0.18
0.39
0.62
0.49
0.31
0.46
0.23
0.42
0.19
0.39
Table 2. Panel A: Means and Standard Deviations. Employed Respondents.
Variable
# of Observations
Age
White
Male
Married
Bachelor Degree
Professional Degree
Respondent Income
(in $1,000 of 1992)
Net Worth
(in $ 1,000)
Non-Risky NW
(in $ 1,000)
Stocks
(in $ 1,000)
Bonds
(in $ 1,000)
Real Estate
(in $ 1,000)
Stock Ownership
Hourly Worker
% Months Worked
Flexible to Increase Hwk.
Flexible to Decrease Hwk.
Second Job
Employed
Sample
15,701
Stock Owners
4,029
Don’t Own
Stocks
11,672
Flexible
Labor Supply
5,185
Second-Job
Sub-Sample
1,722
57.92
5.08
0.71
0.45
0.50
0.50
0.66
0.47
0.26
0.44
0.10
0.30
28.72
32.35
165.50
319.55
33.65
123.73
75.29
260.92
5.38
31.27
93.18
216.20
0.26
0.44
0.56
0.50
0.94
0.20
0.33
0.47
0.32
0.47
0.11
0.31
57.83
5.00
0.89
0.31
0.60
0.49
0.76
0.43
0.43
0.50
0.18
0.39
40.88
47.03
336.69
513.17
64.11
206.57
75.29
260.92
16.07
53.77
119.03
280.28
1.00
0.00
0.40
0.49
0.96
0.17
0.35
0.48
0.34
0.47
0.12
0.33
57.95
5.11
0.65
0.48
0.47
0.50
0.63
0.48
0.20
0.40
0.07
0.26
24.35
23.57
103.35
170.57
23.13
73.73
—
0.00
1.40
14.41
74.24
150.39
0.00
0.00
0.61
0.49
0.94
0.21
0.33
0.47
0.31
0.46
0.10
0.31
58.07
5.34
0.74
0.44
0.54
0.50
0.69
0.46
0.25
0.43
0.08
0.28
29.05
39.92
185.84
382.69
42.57
185.22
79.66
217.85
6.32
36.90
105.59
276.82
0.27
0.44
0.59
0.49
0.94
0.20
1.00
0.00
0.49
0.50
0.11
0.31
56.87
4.96
0.73
0.45
0.55
0.50
0.67
0.47
0.38
0.49
0.16
0.37
33.08
36.60
196.60
358.96
52.47
214.93
56.25
123.90
7.63
34.15
103.37
186.79
0.29
0.45
0.51
0.50
0.95
0.18
0.33
0.47
0.35
0.48
1.00
0.00
25
Table 2. Panel B: Means and Standard Deviations. Employed Respondents.
Variable
# of Observations
Employer Health Insurance
Retiree Health Insurance
Private Health Insurance
No Health Insurance
Have thought about retirement
Prob. Working after 62
Prob. Living to 75
Health Limitation
Excellent Health
Very Good Health
Good Health
Smoker
Drinker
wave1
wave2
wave3
Employed
Sample
15,701
Stock Owners
4,029
Don’t Own
Stocks
11,672
Flexible
Labor Supply
5,185
Second-Job
Sub-Sample
1,722
0.81
0.39
0.79
0.41
0.17
0.37
0.08
0.27
0.48
0.50
0.48
0.40
0.67
0.28
0.18
0.39
0.21
0.40
0.34
0.47
0.31
0.46
0.23
0.42
0.58
0.49
0.27
0.44
0.22
0.41
0.19
0.40
0.83
0.38
0.82
0.39
0.19
0.39
0.02
0.13
0.53
0.50
0.46
0.39
0.69
0.25
0.17
0.38
0.28
0.44
0.41
0.49
0.25
0.43
0.17
0.38
0.73
0.44
0.28
0.45
0.27
0.44
0.18
0.38
0.80
0.40
0.77
0.42
0.16
0.37
0.10
0.30
0.46
0.50
0.49
0.40
0.66
0.29
0.19
0.39
0.18
0.39
0.32
0.47
0.33
0.47
0.25
0.43
0.53
0.50
0.26
0.44
0.20
0.40
0.20
0.40
0.76
0.42
0.80
0.40
0.19
0.39
0.08
0.28
0.46
0.50
0.48
0.40
0.67
0.28
0.19
0.39
0.22
0.41
0.34
0.47
0.31
0.46
0.24
0.43
0.60
0.49
0.32
0.47
0.19
0.39
0.18
0.38
0.81
0.39
0.80
0.40
0.18
0.38
0.07
0.26
0.48
0.50
0.54
0.39
0.69
0.27
0.19
0.39
0.27
0.44
0.37
0.48
0.27
0.45
0.19
0.39
0.60
0.49
0.31
0.46
0.23
0.42
0.19
0.40
26
Table 3: Employed and Self-Employed Respondents. The Logarithm of the stock holdings is the dependent variable
27
No.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
Variable
Constant
White
Male
Married
Bachelor D.
Graduate D.
Age
Age Sq.
Self-Emp. W/Wages
Self-Emp. W/Profits
Second Job
Hourly Worker
Earnings
Non risky wealth
Non risky wlt. Sq.
% Months Worked
No Health Ins.
Private Hlt. Ins.
PWFT62
Health Limitation
Excellent Health
Very Good Health
First Wave
Second Wave
Third Wave
New England
Mid-Atlantic
East N-C.
West N-C.
# Obs.
2
R
OLS Estimates
Estimate Standard Error
7.6428
3.6298
0.3942
0.1322
0.1124
0.0823
0.1094
0.0941
0.5437
0.0939
0.2760
0.1226
-0.0045
0.1324
0.0005
0.0012
0.1479
0.1355
-0.0583
0.1099
-0.2857
0.1041
-0.2604
0.0889
0.0026
0.0004
0.7758
0.0699
-0.0346
0.0043
0.1150
0.2259
-0.4166
0.2877
0.1801
0.0811
-0.5712
0.0918
-0.2161
0.0908
0.1997
0.0942
0.0957
0.0819
-0.5503
0.0907
-0.3379
0.0866
-0.3631
0.0916
0.3547
0.1588
0.1628
0.1123
0.0523
0.1033
0.2360
0.1280
3,967
0.1766
Breusch-Pagan Test Statistic: 480.29. P-value: 0.0000.
Hausman Specification Test Statistic: 119.83. P-value: 0.0000.
Fixed Effects
Estimate Standard Error
8.7862
6.6631
—
—
—
—
-0.0287
0.1850
—
—
—
—
-0.0522
0.2171
0.0013
0.0019
-0.2778
0.1969
-0.0445
0.1503
-0.0140
0.1102
-0.0058
0.1274
0.0009
0.0004
0.2118
0.0712
-0.0092
0.0033
-0.1073
0.1872
0.2648
0.2527
0.2299
0.0867
-0.1175
0.1124
-0.0171
0.0972
-0.0404
0.1174
-0.0155
0.0923
-0.2956
0.3266
-0.1540
0.2321
-0.2094
0.1446
—
—
—
—
—
—
—
—
3,967
0.0543
Random Effects
Estimate Standard Error
8.2947
3.5146
0.3897
0.1296
0.0962
0.0876
0.1198
0.0895
0.5715
0.0993
0.3262
0.1282
-0.0279
0.1277
0.0007
0.0012
0.0583
0.1284
-0.0153
0.0965
-0.1663
0.0855
-0.2646
0.0800
0.0017
0.0003
0.5347
0.0580
-0.0227
0.0028
0.0299
0.1611
-0.1503
0.1775
0.2005
0.0698
-0.4327
0.0801
-0.1403
0.0767
0.0922
0.0837
0.0564
0.0706
-0.5184
0.0878
-0.3232
0.0784
-0.3315
0.0771
0.3124
0.1826
0.2043
0.1250
0.0777
0.1098
0.1992
0.1328
3,967
0.1706
Table 4: Employed and Self-Employed Respondents. An Indicator of stock ownership is the dependent
variable
No.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
Variable
Constant
Thougt about Retirement.
Prob. Living to 75
Smoker
Drinker
Weight
White
Male
Married
Bachelor D.
Graduate D.
Age
Age Sq.
Self-Emp. W/Wages
Self-Emp. W/Profits
Second Job
Hourly Worker
Earnings
Non risky wealth
Non risky wlt. Sq.
% Months Worked
No Health Ins.
Private Hlt. Ins.
PWFT62
Health Limitation
Excellent Health
Very Good Health
First Wave
Second Wave
Third Wave
New England
Mid-Atlantic
East N-C.
West N-C.
# Obs./Avg. Prob./Log Lik.
2
Pseudo-R
Estimate
-0.8332
0.0318
0.0570
-0.1468
0.3178
-0.0009
0.5993
0.1132
0.1269
0.3680
-0.0077
-0.0317
0.0003
-0.0612
0.0391
0.0199
-0.2222
0.0009
0.5412
-0.0228
0.1053
-0.5849
0.0437
-0.1782
-0.0017
0.1977
0.1631
0.0579
0.1691
-0.0332
-0.0552
-0.0388
0.1703
0.0575
13,528
0.1558
Probit
St. Error Marg. Effect
1.6049
—
0.0272
0.0102
0.0591
0.0183
0.0410
-0.0460
0.0349
0.0995
0.0005
-0.0003
0.0477
0.1723
0.0418
0.0364
0.0395
0.0402
0.0465
0.1240
0.0640
-0.0025
0.0590
-0.0102
0.0005
0.0001
0.0659
-0.0193
0.0511
0.0127
0.0446
0.0064
0.0372
-0.0709
0.0005
0.0003
0.0578
0.1740
0.0028
-0.0073
0.0750
0.0339
0.0672
-0.1561
0.0371
0.0142
0.0387
-0.0573
0.0388
-0.0006
0.0414
0.0657
0.0345
0.0533
0.0396
0.0187
0.0369
0.0558
0.0349
-0.0106
0.0809
-0.0174
0.0520
-0.0124
0.0482
0.0568
0.0591
0.0188
0.2557
-6910.76
28
Random Effects Probit
Est.
St. Error Marg. Effect
-0.7476
2.2638
—
0.0662
0.0387
0.0140
0.0940
0.0837
0.0199
-0.2570
0.0607
-0.0505
0.4117
0.0504
0.0829
-0.0013
0.0008
-0.0003
1.0894
0.0775
0.1721
0.1812
0.0655
0.0384
0.2500
0.0608
0.0503
0.7431
0.0758
0.1867
0.0044
0.1020
0.0009
-0.0844
0.0828
-0.0179
0.0008
0.0008
0.0002
-0.0723
0.0943
-0.0148
0.0830
0.0731
0.0182
0.0121
0.0613
0.0026
-0.3125
0.0540
-0.0652
0.0010
0.0004
0.0002
0.6199
0.0612
0.1313
-0.0223
0.0038
-0.0047
0.1735
0.1106
0.0368
-0.7600
0.0975
-0.1113
0.0573
0.0508
0.0124
-0.2394
0.0561
-0.0507
-0.0148
0.0548
-0.0031
0.2715
0.0598
0.0624
0.2235
0.0489
0.0492
0.1398
0.0602
0.0304
0.3315
0.0559
0.0768
-0.0547
0.0546
-0.0114
0.0322
0.1356
0.0069
-0.0222
0.0837
-0.0047
0.3151
0.0773
0.0748
0.1511
0.0931
0.0342
13,528
0.1054
0.1303
-6093.32
Table 5: Selection Corrected Results. Employed and Self-Employed Respondents. The Logarithm of the stock holdings is the dependent variable
29
No.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
Variable
constant
White
Male
Married
Bachelor D.
Graduate D.
Age
Age Sq.
Self-Emp. W/Wages
Self-Emp. W/Profits
Second Job
Hourly Worker
Earnings
Non risky wealth
Non risky wlt. Sq.
% Months Worked
No Health Ins.
Private Hlt. Ins.
PWFT62
Health Limitation
Excellent Health
Very Good Health
First Wave
Second Wave
Third Wave
New England
Mid-Atlantic
East N-C.
West N-C.
Inverse Mills’ Ratio
υit
υi
# Obs.
2
R
Corrected OLS
Estimate Standard Error
10.5117
3.5686
-0.4409
0.1827
-0.0208
0.0831
-0.0708
0.0968
0.0496
0.1221
0.2707
0.1217
0.0204
0.1290
0.0002
0.0012
0.2271
0.1325
-0.1025
0.1100
-0.3042
0.1036
0.0768
0.1002
0.0021
0.0003
0.3079
0.0943
-0.0149
0.0046
0.0180
0.2253
0.4254
0.3084
0.1388
0.0805
-0.3156
0.1020
-0.1964
0.0897
-0.1261
0.1046
-0.1497
0.0887
-0.6516
0.0916
-0.5772
0.0946
-0.3300
0.0908
0.3698
0.1548
0.1722
0.1120
-0.1493
0.1072
0.1520
0.1266
-1.8332
0.2930
—
—
—
—
3,967
0.1882
Cross-Section Corrected RE
Est.
Stand. Error
10.6435
3.5169
-0.3392
0.1742
-0.0277
0.0892
-0.0334
0.0924
0.1336
0.1212
0.3212
0.1273
-0.0013
0.1271
0.0004
0.0012
0.1206
0.1281
-0.0559
0.0962
-0.1853
0.0852
0.0236
0.0921
0.0014
0.0003
0.1545
0.0843
-0.0072
0.0038
-0.0519
0.1611
0.5533
0.2100
0.1606
0.0698
-0.2197
0.0868
-0.1306
0.0764
-0.1655
0.0931
-0.1438
0.0774
-0.6122
0.0887
-0.5346
0.0851
-0.3070
0.0769
0.3167
0.1814
0.2089
0.1242
-0.1033
0.1128
0.1243
0.1324
-1.5899
0.2558
—
—
—
—
3,967
0.1832
Panel Corrected OLS
Est.
Stand. Error
9.6692
3.4819
-0.4196
0.1895
-0.0160
0.0828
-0.0847
0.0979
-0.0396
0.1364
0.2811
0.1202
0.0537
0.1268
-0.0002
0.0012
0.2223
0.1320
-0.0574
0.1086
-0.2874
0.1021
0.0671
0.0988
0.0022
0.0004
0.4060
0.0827
-0.0191
0.0038
0.0319
0.2236
0.3258
0.3072
0.1349
0.0795
-0.3280
0.0986
-0.1957
0.0894
-0.0909
0.1026
-0.1215
0.0876
-0.7953
0.0955
-0.7325
0.1030
-0.3553
0.0916
0.3375
0.1498
0.1491
0.1102
-0.1800
0.1083
0.1131
0.1261
—
—
-1.1500
0.1571
0.3441
0.0518
3,967
0.1997
Panel Corrected RE
Est.
Stand. Error
9.2777
3.4976
-0.2103
0.1820
0.0026
0.0886
-0.0266
0.0937
0.1391
0.1322
0.3335
0.1263
0.0410
0.1270
0.0000
0.0012
0.1057
0.1276
-0.0136
0.0958
-0.1749
0.0851
-0.0115
0.0917
0.0015
0.0003
0.2632
0.0740
-0.0119
0.0033
-0.0543
0.1616
0.4230
0.2037
0.1581
0.0698
-0.2414
0.0857
-0.1342
0.0764
-0.1250
0.0924
-0.1194
0.0770
-0.7283
0.0917
-0.6527
0.0910
-0.3265
0.0774
0.2934
0.1805
0.1841
0.1232
-0.1059
0.1138
0.0999
0.1322
—
—
-0.8581
0.1348
0.2591
0.0513
3,967
0.1931
Table 6: Employed Respondents. The Logarithm of the stock holdings is the dependent variable
30
No.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
Variable
Constant
White
Male
Married
Bachelor D.
Graduate D.
Age
Age Sq.
Flexible Labor Supply
Second Job
Hourly Worker
Earnings
Non risky wealth
Non risky wlt. Sq.
% Months Worked
No Health Ins.
Private Hlt. Ins.
PWFT62
Health Limitation
Excellent Health
Very Good Health
First Wave
Second Wave
Third Wave
New England
Mid-Atlantic
East N-C.
West N-C.
# Obs.
2
R
OLS Estimates
Estimate Standard Error
5.5932
3.9801
0.3430
0.1372
0.0888
0.0944
0.1156
0.1040
0.4629
0.1072
0.3154
0.1330
0.0786
0.1459
-0.0003
0.0013
0.1688
0.0796
-0.3347
0.1292
-0.2076
0.0956
0.0062
0.0011
1.0679
0.0963
0.0000
0.0000
-0.0754
0.2291
0.0425
0.3586
-0.0752
0.0968
-0.5662
0.1032
-0.1788
0.1053
0.1348
0.1105
0.0741
0.0935
-0.5424
0.1103
-0.3207
0.1066
-0.3878
0.1165
0.2777
0.1797
0.0890
0.1244
-0.0244
0.1122
0.1394
0.1490
2,769
0.1638
Breusch-Pagan Test Statistic: 210.45. P-value: 0.0000.
Hausman Specification Test Statistic: 90.29. P-value: 0.0000.
Fixed Effects
Estimate Standard Error
7.8838
8.6100
—
—
—
—
-0.1072
0.2442
—
—
—
—
0.0655
0.2831
-0.0003
0.0025
0.1457
0.0892
-0.0453
0.1568
-0.0380
0.1588
-0.0004
0.0016
0.3497
0.2231
0.0000
0.0000
-0.2912
0.2589
1.0986
0.5809
-0.1876
0.1091
-0.0775
0.1473
-0.1433
0.1262
-0.1445
0.1522
-0.0747
0.1187
-0.6342
0.4347
-0.3372
0.3126
-0.3326
0.1938
—
—
—
—
—
—
—
—
2,769
0.0031
Random Effects
Estimate Standard Error
5.7022
4.0135
0.3671
0.1397
0.0780
0.0990
0.1287
0.1028
0.4783
0.1118
0.3937
0.1429
0.0816
0.1465
-0.0003
0.0013
0.1698
0.0700
-0.2299
0.1069
-0.2323
0.0887
0.0044
0.0010
0.7665
0.0976
0.0000
0.0000
-0.2020
0.2087
0.3201
0.3739
-0.1024
0.0884
-0.4809
0.0958
-0.1831
0.0931
0.0544
0.1000
0.0569
0.0842
-0.5209
0.1081
-0.2897
0.0992
-0.3330
0.0992
0.2277
0.2144
0.1454
0.1381
-0.0452
0.1209
0.1176
0.1473
2,769
0.1592
Table 7: Employed Respondents. An Indicator of stock ownership is the dependent variable
No.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
Variable
Constant
Thought about Retirement.
Prob. Living to 75
Smoker
Drinker
Weight
White
Male
Married
Bachelor D.
Graduate D.
Age
Age Sq.
Flexible Labor Supply
Second Job
Hourly Worker
Earnings
Non risky wealth
Non risky wlt. Sq.
% Months Worked
No Health Ins.
Retiree Health Ins.
PWFT62
Health Limitation
Excellent Health
Very Good Health
First Wave
Second Wave
Third Wave
New England
Mid-Atlantic
East N-C.
West N-C.
Estimate
-1.4556
0.0254
0.0924
-0.1248
0.3081
-0.0011
0.5702
0.0739
0.1228
0.3112
-0.0348
-0.0101
0.0001
0.0396
0.0046
-0.1815
0.0045
0.6271
-0.0334
0.0789
-0.5290
0.0905
-0.2052
0.0549
0.1378
0.1349
0.0455
0.1415
-0.0321
-0.0789
-0.0803
0.1318
0.1024
# Obs./Avg. Prob/Log Lik.
2
Pseudo-R
9,274
0.1425
Probit
St. Error Marg. Effect
1.9831
—
0.0329
0.0083
0.0692
0.0304
0.0472
-0.0402
0.0407
0.0990
0.0006
-0.0004
0.0539
0.1695
0.0517
0.0243
0.0455
0.0398
0.0550
0.1062
0.0754
-0.0114
0.0726
-0.0033
0.0007
0.0000
0.0348
0.0131
0.0536
0.0015
0.0434
-0.0598
0.0015
0.0015
0.0772
0.2063
0.0043
-0.0110
0.0954
0.0260
0.1200
-0.1441
0.0418
0.0293
0.0456
-0.0675
0.0455
0.0183
0.0490
0.0464
0.0401
0.0449
0.0482
0.0150
0.0461
0.0473
0.0439
-0.0105
0.0929
-0.0254
0.0605
-0.0259
0.0547
0.0445
0.0693
0.0345
0.2673
-4848.13
31
Random Effects Probit
Est.
St. Error Marg. Effect
-2.3298
2.7019
—
0.0813
0.0466
0.0184
0.1340
0.0992
0.0304
-0.2468
0.0715
-0.0523
0.4087
0.0592
0.0885
-0.0018
0.0010
-0.0004
1.0278
0.0871
0.1787
0.1062
0.0760
0.0241
0.2459
0.0713
0.0533
0.6409
0.0870
0.1664
-0.0268
0.1157
-0.0060
-0.0254
0.0987
-0.0058
0.0003
0.0009
0.0001
0.0359
0.0497
0.0082
0.0071
0.0758
0.0016
-0.2645
0.0614
-0.0605
0.0057
0.0009
0.0013
0.7611
0.0856
0.1729
-0.0406
0.0059
-0.0092
0.2275
0.1424
0.0517
-0.8398
0.2065
-0.1196
0.0788
0.0597
0.0175
-0.2891
0.0673
-0.0657
0.0596
0.0658
0.0138
0.2141
0.0721
0.0517
0.2028
0.0582
0.0475
0.1005
0.0739
0.0233
0.2942
0.0692
0.0709
-0.0525
0.0685
-0.0117
-0.0369
0.1549
-0.0082
-0.1046
0.0959
-0.0228
0.2611
0.0880
0.0648
0.1999
0.1083
0.0493
9.274
0.1018
0.1442
-4334.30
Table 8: Selection Corrected Results. Employed Respondents. The Logarithm of the stock holdings is the dependent variable
32
No.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
Variable
Constant
White
Male
Married
Bachelor D.
Graduate D.
Age
Age Sq.
Flexible Labor Supply
Second Job
Hourly Worker
Earnings
Non risky wealth
Non risky wlt. Sq.
% Months Worked
No Health Ins.
Retiree Health Ins.
PWFT62
Health Limitation
Excellent Health
Very Good Health
First Wave
Second Wave
Third Wave
New England
Mid-Atlantic
East N-C.
West N-C.
Inverse Mills’ Ratio
υit
υi
# Obs.
2
R
Corrected OLS
Estimate St. Error
9.2477
3.9430
-0.4402
0.1939
0.0044
0.0935
-0.0336
0.1045
0.0404
0.1334
0.3300
0.1323
0.0748
0.1430
-0.0003
0.0013
0.1158
0.0795
-0.3363
0.1278
0.0723
0.1049
0.0023
0.0012
0.4116
0.1591
-0.0219
0.0078
-0.1591
0.2287
0.7598
0.3494
-0.1782
0.0996
-0.2892
0.1146
-0.2197
0.1041
-0.1179
0.1172
-0.1260
0.0989
-0.6304
0.1099
-0.5325
0.1108
-0.3681
0.1157
0.2982
0.1757
0.1414
0.1249
-0.2070
0.1149
0.0065
0.1493
-1.8044
0.3384
—
—
—
—
2,769
0.1745
Cross-Section Corrected RE
Estimate
Stand. Error
8.7456
4.0363
-0.3646
0.1984
-0.0092
0.0998
-0.0108
0.1057
0.0794
0.1353
0.4087
0.1420
0.0914
0.1457
-0.0004
0.0013
0.1216
0.0702
-0.2345
0.1064
0.0289
0.1017
0.0010
0.0011
0.1488
0.1545
-0.0086
0.0082
-0.2786
0.2084
0.9756
0.3934
-0.2079
0.0904
-0.2320
0.1069
-0.2295
0.0931
-0.1650
0.1082
-0.1219
0.0907
-0.6057
0.1088
-0.4865
0.1058
-0.3136
0.0989
0.2332
0.2130
0.1919
0.1374
-0.2189
0.1247
-0.0147
0.1485
-1.6875
0.3272
—
—
—
—
2,769
0.1701
Panel Corrected OLS
Est.
St. Error
8.7687
3.9060
-0.4084
0.1994
0.0185
0.0931
-0.0598
0.1054
-0.0458
0.1464
0.3186
0.1315
0.0878
0.1425
-0.0005
0.0013
0.1308
0.0784
-0.3232
0.1274
0.0511
0.1035
0.0035
0.0012
0.5425
0.1387
-2.93e-07 6.75e-08
-0.2366
0.2331
0.7049
0.3657
-0.1557
0.0969
-0.3063
0.1123
-0.2025
0.1037
-0.1024
0.1170
-0.1012
0.0979
-0.7045
0.1136
-0.6352
0.1181
-0.3746
0.1161
0.2801
0.1738
0.1307
0.1234
-0.2236
0.1158
-0.0119
0.1488
—
—
-1.0483
0.1768
0.2443
0.0579
2,769
0.1803
Panel Corr. RE
Est.
St. Error
7.9463
4.0283
-0.2628
0.2029
0.0204
0.0992
-0.0172
0.1076
0.0598
0.1458
0.3962
0.1416
0.1061
0.1458
-0.0006
0.0013
0.1417
0.0699
-0.2311
0.1064
-0.0135
0.1001
0.0022
0.0011
0.3156
0.1350
-1.75e-07 7.33e-08
-0.3497
0.2109
0.9165
0.3977
-0.1755
0.0893
-0.2657
0.1055
-0.2162
0.0929
-0.1369
0.1085
-0.0973
0.0907
-0.6569
0.1114
-0.5605
0.1113
-0.3187
0.0995
0.2025
0.2129
0.1750
0.1370
-0.2191
0.1263
-0.0228
0.1490
—
—
-0.8408
0.1661
0.1588
0.0579
2,769
0.1746
Mdlk. Panel Corr. RE
Est.
St. Error
8.6242
3.9515
-0.5571
0.1711
-0.0043
0.0971
-0.1018
0.1036
-0.1464
0.1278
0.3884
0.1395
0.1357
0.1439
-0.0009
0.0013
0.1224
0.0690
-0.2304
0.1050
0.0491
0.0921
0.0014
0.0010
0.0281
0.1250
5.41e-07 1.19e-07
-0.4694
0.2074
1.0770
0.3783
-0.1887
0.0875
-0.1917
0.0988
-0.2108
0.0916
-0.1924
0.1022
-0.1299
0.0856
-0.7171
0.1095
-0.7055
0.1063
-0.3307
0.0981
0.2063
0.2096
0.1865
0.1349
-0.3060
0.1215
-0.0832
0.1453
—
—
-1.2700
0.1262
0.2484
0.0567
2,769
0.1942
Table A.1.: Sensitivity Analysis. Employed and Self-Employed Respondents. The Logarithm of the stock holdings is the dependent variable, except in
columns (3) and (4).
33
No.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
Variable
Constant
White
Male
Married
Bachelor D.
Graduate D.
Age
Age Sq.
Self-Emp. Indicator
Self-Emp. w/Wages
Self-Emp. w/Profits
Second Job
Hourly Worker
Earnings
Non risky wealth
Non risky wlt. Sq.
% Months Worked
No Health Ins.
Private Hlt. Ins.
PWFT62
PWFT65
Health Limitation
Excellent Health
Very Good Health
First Wave
Second Wave
Third Wave
New England
Mid-Atlantic
East N-C.
West N-C.
υit
υi
# Obs.
2
R
Estimate
9.2959
-0.2076
0.0058
-0.0239
0.1411
0.3298
0.0401
0.0000
0.0059
—
—
-0.1760
-0.0199
0.0015
0.2682
-0.0121
-0.0539
0.4253
0.1599
-0.2416
—
-0.1338
-0.1245
-0.1186
-0.7300
-0.6526
-0.3267
0.2990
0.1867
-0.1031
0.1021
-0.8544
0.2589
3,967
0.1927
(1)
Standard Error
3.4980
0.1820
0.0885
0.0936
0.1327
0.1263
0.1270
0.0012
0.0951
—
—
0.0851
0.0916
0.0003
0.0735
0.0033
0.1617
0.2051
0.0697
0.0858
—
0.0764
0.0925
0.0770
0.0917
0.0911
0.0774
0.1804
0.1232
0.1137
0.1322
0.1350
0.0514
Est.
9.4193
-0.3540
-0.0218
-0.0089
0.0520
0.2711
0.0551
-0.0002
—
0.1620
0.0363
-0.2193
-0.0189
0.0010
0.2662
-0.0112
-0.2054
0.4584
0.1313
—
-0.1725
-0.1638
-0.0892
-0.1449
-0.7318
-0.6546
-0.3955
0.3687
0.2306
-0.1427
0.1069
-1.0022
0.2959
3,394
0.1877
(2)
Stand. Error
3.5675
0.1862
0.0938
0.0992
0.1381
0.1303
0.1288
0.0012
—
0.1373
0.1028
0.0940
0.1046
0.0003
0.0806
0.0036
0.2050
0.2233
0.0776
—
0.1113
0.0855
0.1021
0.0865
0.0980
0.0995
0.0895
0.1881
0.1286
0.1187
0.1390
0.1488
0.0551
Est.
6.3389
-0.1589
-0.0579
0.1828
0.1077
0.3267
0.1501
-0.0011
—
0.2293
0.3683
0.0690
-0.0949
0.0012
0.3338
-0.0136
-0.1230
0.3041
0.1275
-0.1937
—
-0.1867
-0.0543
-0.0761
-0.5415
-0.4236
-0.1962
-0.0152
0.0729
-0.1651
0.0542
-0.8415
0.4091
5,801
0.2096
(3)
Stand. Error
2.7647
0.1305
0.0710
0.0829
0.1018
0.1079
0.0992
0.0009
—
0.0966
0.0741
0.0654
0.0733
0.0003
0.0611
0.0028
0.1221
0.1404
0.0547
0.0665
—
0.0590
0.0698
0.0557
0.0737
0.0723
0.0597
0.1535
0.1021
0.0902
0.1065
0.1254
0.0449
Est.
11.6458
-0.1428
0.0258
-0.0125
0.1343
0.3813
-0.0458
0.0008
—
0.1025
-0.0175
-0.1340
-0.0278
0.0015
0.2649
-0.0116
-0.0431
0.4662
0.1951
-0.2546
—
-0.1507
-0.0679
-0.0981
-0.8132
-0.6951
-0.3651
0.2069
0.0976
-0.1347
0.0613
-0.9105
0.3088
4,165
0.2192
(4)
Stand. Error
3.4333
0.1768
0.0863
0.0907
0.1289
0.1246
0.1234
0.0011
—
0.1230
0.0924
0.0817
0.0877
0.0003
0.0702
0.0032
0.1534
0.1950
0.0668
0.0821
—
0.0732
0.0869
0.0721
0.0878
0.0858
0.0736
0.1776
0.1207
0.1122
0.1296
0.1294
0.0507
Table A.2.: Sensitivity Analysis. Employed Respondents. The Logarithm of the stock holdings is the dependent variable, except in columns (3) and (4).
34
No.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
Variable
Constant
White
Male
Married
Bachelor D.
Graduate D.
Age
Age Sq.
Labor Supply Flexibility
Second Job
Hourly Worker
Earnings
Non risky Wealth
Non risky Wlt. Sq.
% Months Worked
No Health Ins.
Retiree Health Ins.
PWFT62
PWFT65
Health Limitation
Excellent Health
Very Good Health
First Wave
Second Wave
Third Wave
New England
Mid-Atlantic
East N-C.
West N-C.
υit
υi
# Obs.
2
R
Estimate
7.7348
-0.3084
0.0125
0.0285
0.0294
0.3106
0.1225
-0.0008
0.1606
-0.2859
-0.0979
0.0031
0.4725
-0.0255
-0.6314
0.5461
-0.2465
—
-0.2360
-0.2292
-0.0502
-0.0601
-0.6396
-0.4858
-0.4006
0.2417
0.1991
-0.2494
-0.0162
-0.8254
0.2304
2,314
0.1727
(1)
Standard Error
4.1814
0.2166
0.1085
0.1168
0.1612
0.1492
0.1493
0.0014
0.0782
0.1180
0.1156
0.0013
0.1467
0.0081
0.2763
0.6659
0.1017
—
0.1448
0.1050
0.1206
0.1029
0.1224
0.1246
0.1161
0.2260
0.1470
0.1357
0.1630
0.1931
0.0648
(2)
Est.
St. Error
8.5397
4.0613
-0.2782
0.2004
0.0419
0.1006
-0.0079
0.1100
0.0588
0.1458
0.4046
0.1441
0.0924
0.1472
-0.0005
0.0013
0.1462
0.0702
-0.2442
0.1075
-0.0192
0.1011
0.0012
0.0009
0.0341
0.0405
-0.0003
0.0010
-0.3925
0.2111
1.0748
0.3984
-0.1769
0.0892
-0.2730
0.1066
—
—
-0.2340
0.0932
-0.1399
0.1092
-0.1019
0.0916
-0.6801
0.1125
-0.5763
0.1122
-0.3238
0.0995
0.1560
0.2155
0.1782
0.1394
-0.2459
0.1266
-0.0628
0.1504
-0.8822
0.1550
0.1918
0.0580
2,769
0.1541
(3)
Est.
St. Error
4.1925
3.2748
-0.1752
0.1506
-0.1398
0.0812
0.2197
0.0979
0.1199
0.1255
0.3521
0.1207
0.2338
0.1169
-0.0018
0.0011
0.1057
0.0546
0.0470
0.0821
-0.1068
0.0789
0.0029
0.0009
0.4321
0.1107
-0.0239
0.0062
-0.2974
0.1587
0.1834
0.3023
-0.0590
0.0672
-0.2159
0.0866
—
—
-0.2029
0.0723
-0.0407
0.0838
-0.0546
0.0674
-0.6381
0.0924
-0.4553
0.0924
-0.2513
0.0764
-0.0842
0.1824
0.0985
0.1170
-0.2420
0.1007
-0.0016
0.1218
-0.7958
0.1708
0.3772
0.0509
3,930
0.1723
(4)
Est.
St. Error
9.5902
3.9447
-0.1782
0.1938
0.0107
0.0965
-0.0051
0.1032
0.0455
0.1393
0.3938
0.1389
0.0440
0.1414
0.0000
0.0013
0.1298
0.0664
-0.2114
0.1013
-0.0657
0.0947
0.0024
0.0011
0.3300
0.1273
-0.0184
0.0070
-0.2337
0.1979
0.7685
0.3785
-0.1990
0.0851
-0.2798
0.0996
—
—
-0.2300
0.0883
-0.0767
0.1014
-0.0874
0.0843
-0.7813
0.1068
-0.6179
0.1047
-0.4089
0.0934
0.1526
0.2089
0.1004
0.1334
-0.2584
0.1235
-0.0612
0.1454
-0.9004
0.1578
0.2109
0.0569
2,902
0.2001
References
Alessie, R., A. Lusardi, and A. Kapteyn (1999): “Saving After Retirement: Evidence from Three Different
Surveys,” Labour Economics, 6-2 277–310.
Alessie, R., S. Hochguertel, and A. van Soest (2002): “Household Portfolios in the Netherlands,” in Household Portfolios, by L. Guiso, M. Haliassos, and T. Jappelli (eds.) MIT Press.
Ameriks, J., and S.P. Zeldes (2001): “How do Household Portfolio Shares Vary with Age?” manuscript,
TIAA-CREF and Columbia University GSB.
Arrondel, L. (2000): “Risk Management, liquidity constraints, and wealth accumulation behavior in
France.” manuscript. CNRS-DELTA Paris.
Banks, J., R. Blundell, and S. Tanner (1998): “Is There a Retirement-Savings Puzzle?” American Economic
Review, 88-4 769–788.
Benı́tez-Silva, H. (2002): “The Annuity Puzzle Revisited,” manuscript, SUNY-Stony Brook.
Benı́tez-Silva, H., and D.S. Dwyer (2003): “What to Expect when you are Expecting Rationality: Testing
Rational Expectations using Micro Data,” manuscript, SUNY-Stony Brook.
Bernheim, B.D. (1987): “Dissaving after Retirement: Testing the Pure Life Cycle Hypothesis,” in Z. Bodie,
J.B. Shoven, and D.A. Wise (eds.) Issues in Pension Economics, 237–274.
Bertaut, C., and M. Starr-McCluer (2002): “Household Portfolios in the United States,” in Household
Portfolios, by L. Guiso, M. Haliassos, and T. Jappelli (eds.) MIT Press.
Blundell, R., and C. Meghir (1987): “Bivariate Alternatives to the Tobit Model,” Journal of Econometrics,
34 179–200.
Bodie, Z., and W.F. Samuelson (1989): “Labor Supply Flexibility and Portfolio Choice,” NBER Working
Paper Series, 3043.
Bodie, Z., R.C. Merton, and W.F. Samuelson (1992): “Labor supply flexibility and portfolio choice in a
life cycle model,” Journal of Economic Dynamics and Control, 16 427–449.
Campbell, J.Y., J. Cocco, F. Gomes, P.J. Maenhout, and L.M. Viceira (2001): “Stock Market Mean Reversion and the Optimal Equity Allocation of a Long Lived Investor,” manuscript, Harvard University.
Campbell, J.Y., and J. Cocco (2002): “Household Risk Management and Optimal Mortgage Choice,”
manuscript, Harvard University.
Cochrane, J.H. (1997): “Where is the Market Going? Uncertain Facts and Novel Theories,” NBER Working
Paper Series, 6207.
Coronado, J.L., and M. Perozek (2001): “Wealth Effects and the Consumption of Leisure: Retirement
Decisions During the Stock Market Boom of the 1990s,” manuscript, Federal Reserve Board.
D’Addio, A.C., De Greef, I., and M. Rosholm (2002): “Assessing Unemployment Traps in Belgium using
Panel Data Sample Selection Models,” Discussion Paper 669. IZA Bonn.
Davies, J.B. (1981): “Uncertain Lifetime, Consumption, and Dissaving in Retirement,” Journal of Political
Economy, 89-3 561–577.
35
Davis, S.J., and P. Willen (2000): “Uncertain Labor Income and Portfolio Choice: Covariance and Its
Implications,” manuscript, University of Chicago and Princeton University.
Deaton, A., (1997): The Analysis of Household Surveys. Baltimore: Johns Hopkins University Press.
Dustmann. C. and M.E. Rochina-Barrachina (2000): “Selection Correction in Panel Data Models: An
Application to Labour Supply and Wages,” Discussion Paper 162. IZA Bonn.
French, E. (2000): “The Effects of Health, Wealth, and Wages on Labor Supply and Retirement Behavior,”
manuscript. University of Wisconsin and Federal Reserve Bank of Chicago.
Flodén, M. (1998): “Labor Supply and Consumption under Uncertainty: Prudence Reconsidered,”
manuscript, Stockholm University.
Fratantoni, MC (1998): “Homeownership and investment in risky assets,” Journal of Urban Economics,
44 27–42.
Friedman, B.M., and M.J. Warshawsky (1990): “The Cost of Annuities: Implications for Saving Behavior
and Bequests,” Quarterly Journal of Economics, 105-1 135–154.
Gourieroux, C., A. Monfort, E. Renault, A. Trognon (1987): “Generalized Residuals,” Journal of Econometrics, 34 5–32.
Guiso, L., T. Jappelli, and D. Terlizzese (1996): “Income Risk, Borrowing Constraints, and Portfolio
Choice,” American Economic Review, 86-1 158–172.
Guiso, L., M. Haliassos, and T. Jappelli (2002): Household Portfolios. MIT Press.
Gustman, A. L., O. Mitchell, and T.L. Steinmeier (1994): “Retirement Research Using the Health and
Retirement Survey,” NBER Working Paper No. 4813.
Gustman, A. L., O. Mitchell, and T.L. Steinmeier (1995): “Retirement Measures in the Health and Retirement Study,” The Journal of Human Resources 30 (Supplement).
Heaton, J., and D. Lucas (2000a): “Portfolio Choice in the Presence of Background Risk,” The Economic
Journal, 110-1 1–26.
Heaton, J., and D. Lucas (2000b): “Portfolio Choice and Asset Prices: The Importance of Entrepreneurial
Risk,” The Journal of Finance, 55-3 1163–1198.
Heckman, J.J. (1974): “Life Cycle Consumption and Labor Supply: an Explanation of the Relationship
between Income and Consumption over the Life Cycle,” American Economic Review, 64-1 188–194.
Heckman, J.J. (1979): “Sample Selection Bias as a Specification Error,” Econometrica, 47-1 153–161.
Hewett, P.C., and M.R. Montgomery (2003): “Measuring Living Standards with Demographic Surveys: A
Structural Equations Approach to Urban Reproductive Health,” manuscript. Population Council and
SUNY-Stony Brook.
Hochguertel, S. (2003): “Precautionary Motives and Portfolio Decisions,” Journal of Applied Econometrics, 18 61-77.
Hsiao, C. (1986): Analysis of Panel Data. Econometric Society Monographs. Cambridge University Press.
Hurd, M., and M. Reti (2001): “The Effects of Large Capital Gains on Work and Consumption: Evidence
from four waves of the HRS,” manuscript, RAND Corporation: Santa Monica.
36
Hurd, M. (2002): “Portfolio Holdings of the Elderly,” in Household Portfolios, by L. Guiso, M. Haliassos,
and T. Jappelli (eds.) MIT Press.
Jensen, P., M. Rosholm, and M. Verner (2001): “A Comparison of Different Estimators for Panel Data
Sample Selection Models,” Working Paper 2002-1. Department of Economics Aarhus School of
Business and University of Aarhus.
Juster, T.F., and R. Suzman (1995): “An Overview of the Health and Retirement Study,” The Journal of
Human Resources, 30 (Supplement): S7–S56.
Killingsworth, M.R. (1983): Labor Supply. Cambridge University Press. Cambridge, MA.
Kocherlakota, N. (1996): “The Equity Premium: It’s Still a Puzzle,” Journal of Economic Literature, 34-1
42–71.
Koo, H.K. (1999): “Consumption and portfolio selection with labor income: A discrete-time approach,”
Mathematical Methods of Operations Research, 50 219–243.
Kyriazidou, E. (1997): “Estimation of a Panel Data Sample Selection Model,” Econometrica, 65-6 1335–
1364.
Kyriazidou, E. (2001): “Estimation of Dynamic Panel Data Sample Selection Models,”Review of Economic
Studies, 68-3 543–572.
Low, H.W. (1998): “Simulation Methods and Economic Analysis,” manuscript, University College London, University of London.
Low, H.W. (1999): “Self Insurance and Unemployment Benefit in a Life-Cycle Model of Labour Supply
and Savings,” manuscript, University of Cambridge.
Lusardi, A. (1999): “Information, Expectations, and Savings for Retirement,” Behavioral Dimensions
of Retirement Economics. Henry J. Aaron, editor. Brookings Institution Press and Russell Sage
Foundation. Washington D.C. and New York.
Maddala, G.S. (1988): Introduction to Econometrics. Macmillan Publishing Company: New York.
Merton, R.C. (1969): “Lifetime Portfolio Selection Under Uncertainty: the Continuous-Time Case,” Review of Economics and Statistics, 51-3 247–257.
Merton, R.C. (1990): Continuous-Time Finance. Blackwell Publishers. Cambridge, MA.
Miniaci, R., and G. Weber (2002): “Econometric Issues in the Estimation of Household Portfolio Models,”
in Household Portfolios, by L. Guiso, M. Haliassos, and T. Jappelli (eds.) MIT Press.
Mitchell, O.S., J.M. Poterba, M.J. Warshawsky, and J.R. Brown (1999): “New Evidence on the Money’s
Worth of Individual Annuities,” American Economic Review, 89-5 1299-1318.
Moore, J.F., and O.S. Mitchell (2000): “Projected Retirement Wealth and Saving Adequacy,” in Forecasting Retirement Needs and Retirement Wealth, O.S. Mitchell, P.B. Hammond, and A.M. Rappaport
(eds.) University of Pennsylvania Press, Philadelphia.
Nijman, T., and M. Verbeek (1992): “Nonresponse in Panel Data: The Impact on Estimates of a Life Cycle
Consumption Function,” Journal of Applied Econometrics, 7 243–257.
Poterba, J.M., and A.A. Samwick (1997): “Household Portfolio Allocation over the Life Cycle,” NBER
Working Paper Series, 6185.
37
Robst, J., R. Deitz, and K. McGoldrick (1999): “Income variability, uncertainty, and housing tenure
choice,” Regional Science and Urban Economics, 29 219–229.
Rosen, H.S., and S. Wu (2003): “Portfolio Choice and Health Status,” NBER Working Paper Series, 9453.
Rust, J., M. Buchinsky, and H. Benı́tez-Silva (2001): “Dynamic Structural Models of Retirement and
Disability,” manuscript. Presented at the ASSA 2001 Meetings of the Econometric Society.
Svensson, L.E.O. (1988): “Portfolio Choice and Asset Pricing With Nontraded Assets,” NBER Working
Paper Series, 2774.
Svensson, L.E.O., and I. Werner (1993): “Nontraded Assets in incomplete markets: Pricing and Portfolio
Choice,” European Economic Review, 37-5 1149–1168.
Tracy, J., H. Schneider, and S. Chan (1999): “Are Stocks Overtaking Real State in Household Portfolios?”
Current Issues in Economics and Finance. Federal Reserve Bank of New York. 5-5 1–6.
Tracy, J. and H. Schneider (2001): “Stocks in the Household Portfolio: A Look Back at the 1990s.” Current
Issues in Economics and Finance. Federal Reserve Bank of New York. 7-4 1–6.
Vella, F. (1998): “Estimating Models with Sample Selection Bias: A Survey,” Journal of Human Resources,
33-1 127–169.
Vella, F., and M. Verbeek (1999): “Two-step estimation of panel data models with censored endogenous
variables and selection bias,” Journal of Econometrics, 90 239–263.
Verbeek, M., and T. Nijman (1992): “Testing for Selectivity Bias in Panel Data Models,” International
Economic Review, 33 681–703.
Verbeek, M., and T. Nijman (1996): “Incomplete Panels and Selection Bias,” in The Econometrics of Panel
Data, edited by L. Mátyás and P. Sevestre. Klüver Academic Publishers: New York.
Vissing-Jorgenson, A. (1999): “Towards an explanation of household portfolio choice heterogeneity: nonfinancial income and participation cost structures,” manuscript, University of Chicago.
Wooldridge, J.M. (2002): Econometric Analysis of Cross Section and Panel Data. MIT Press.
Yaari, M.E. (1965): “Uncertain Lifetime, Life Insurance, and the Theory of the Consumer,” Review of
Economic Studies, 32-2 137–150.
38
Download